Add ESC-50 audio classification recipe for ESPnet3 - #6762

pytest names a test module outside a package after its basename alone, so the
two test_metrics.py files collide and collection aborts for the whole espnet3
suite, not just for the colliding files. Every other directory under
test/espnet3/ already carries an __init__.py; asr/metrics and cls/metrics did
not, and asr/metrics only got away with it while it had no namesake.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

save_espnet_config used rstrip("\n") where every task's build_model uses
rstrip(), so a token list with trailing whitespace would have been read two
different ways. Also route the warning through a module logger, as the rest of
espnet3/utils does.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`BaseSystem` is for what every system has. `remove_long_short` is used by
TTS and CLS only, so it does not belong there; the earlier move to
`systems/base/` is reverted and the provider, the runner and their test
are copied into `systems/cls/` instead.

TTS returns to its state on master: this branch no longer touches it.

The CLS copy differs from the TTS one by one comment, which now points at
`cls.sh` rather than `tts.sh`. Both scripts apply the same strict-inequality
awk filter, so the behaviour is unchanged.

Review feedback:

- make the MELD builder's module-internal functions private; `MELDBuilder`
  and `resolve_data_root` stay public because `dataset/__init__.py` and
  `dataset/dataset.py` import them
- spell the catch-all parameter `**kwargs`, matching the `DatasetBuilder`
  protocol
- give the MELD demo its own description rather than inheriting the
  TEMPLATE default of `readme.md`, which here is the recipe readme -
  download sizes and training commands, not something a demo visitor needs
- add `remove_long_short` and `prepare_labels` to `pretrain_stages` in
  `egs3/TEMPLATE/cls/run.py`, so running either without `--training_config`
  fails with the guardrail message instead of a later, vaguer error. The
  comment named `prepare_labels`; `remove_long_short` had the same gap
- write the metric config examples as `.. code-block:: yaml` so Sphinx
  highlights them, and add one to `wa.py`, `ua.py` and `macro_f1.py`
- add Examples to every function in `metrics/scoring_utils.py`
- relax `max_wav_duration` from 20 s to 30 s. On MELD this returns six
  training and two validation utterances; the reported numbers still come
  from the 20 s run and are refreshed separately

Found while writing the unit tests:

- `build_model` resolved the model class with `getattr(args, "model", None)`,
  and `ClassChoices.get_class(None)` returns `None` rather than raising, so a
  config without a `model` key died with `'NoneType' object is not callable`.
  The default is now `"espnet"`, matching `ASRTask`, while an explicit
  `model: null` still fails like `encoder: null` does
- the `build_model` return annotation was narrowed to the ESPnet3 subclass,
  which would reject any future entry in `model_choices` that subclasses the
  ESPnet2 model directly. It is back to `ESPnetClassificationModel`
- `freeze_param` naming a parameter that does not exist only warned, and
  training then ran with nothing frozen - the failure the subclass exists to
  prevent. It now raises and lists the available modules
- `AUC` returned `nan` when no class had a negative reference, which reaches
  `metrics.json` as invalid JSON with nothing reporting it. Such classes are
  now skipped, and an input where none qualifies raises
- `prepare_labels` silently ignored rows carrying no label. It now counts and
  reports them, as `remove_long_short` already does for its own drops

`espnet3/systems/cls/{task,espnet_model,system}.py` reach 100% line
coverage, and the metric implementations 97-100%.

test_task.py has three parts: the tests from
`test/espnet2/tasks/test_cls.py`, kept so the copied task cannot drift from
its original unnoticed; one test per ESPnet3 change; and the `build_model`
branches neither file covers. Reverting any of the three changes fails only
the tests written for it.

test_espnet_model.py covers the freeze itself, including that a name matches
only itself and its subtree, and that the freeze survives a backward pass.

test_system.py mirrors `test/espnet3/systems/tts/test_system.py` for
`remove_long_short`, and covers what `prepare_labels` does differently from
`create_token_list`: whitespace-separated labels, rows without one, an empty
manifest, and labels reaching the token list unchanged - the reference labels
are read from the same column, so a label rewritten only there would no
longer match them.

test_metrics.py gains a check that WA / mAP / AUC still agree with
`cls_score.py` to two decimal places on the same inputs, which the pull
request description claimed without anything verifying it.

`ci/test_integration_espnet3.sh` covered ASR only, so nothing exercised the
CLS stages end to end. Adds `egs3/mini_an4/cls` and runs it from
`create_dataset` through `measure` in about four seconds: seven utterances,
log-mel features, a two-block transformer, one epoch on one batch. No
upstream model is downloaded, so CI does not reach the network.

The class label is the AN4 speaker id, following `egs2/mini_an4/lid1`, which
uses it as the language id; AN4 publishes no other per-utterance attribute.
AN4's test speakers never appear in training, so `dataset/config.yaml`
relabels them from a fixed table. The validation utterances are named rather
than counted, because any contiguous pair of the five training utterances
would leave validation with a single class.

The AN4 archive is read from `../asr/downloads.tar.gz` rather than copied.

Also adds the `cls` extras group: `pip install espnet` could not build a CLS
model, because `espnet2/cls/espnet_model.py` raises ImportError without
`torcheval`, which was declared nowhere, and the espnet3 classification
metrics need scikit-learn. `espnet-s3prl` is included because SSL frontends
are the usual choice for classification and it was only reachable through the
s2st extras.

`egs3/mini_an4/cls/dataset/` needs `git add -f`: `egs*/*/*/data*` in
`.gitignore` matches it.

Initialize the parallel runtime for create_dataset stages and add an
audio conversion provider and runner for distributed ffmpeg execution.

Collect MELD conversion jobs before writing manifests, skip existing WAV
files, and use atomic renames to avoid preserving partial outputs.

egs*/*/*/data* also matched egs3/<recipe>/<task>/dataset/, so `git add`
silently skipped every recipe's dataset module. Match `data` and `data_*`
instead, which still covers data/ and data_*/ outputs.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The espnet3 dev-setup guide builds the environment with pixi, which leaves
a .pixi/ directory in the checkout. Keep .pixi/config.toml, which is the
part worth sharing.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Ports egs2/esc50/asr1 to ESPnet3's CLSTask with a BEATs encoder and a
linear head. egs2 expressed 50-way classification through asr.sh; this
uses the real classification system instead.

Reproduces egs2's 5-fold protocol: ESC50_FOLD holds one fold out as both
valid and test, the other four train. The resampled 16 kHz audio does not
depend on the fold, so all five share one wav/ directory.

The corpus is only ever read, so a read-only mount works: prepare_source
validates and reports where to put ESC-50 rather than downloading it.

5-fold average test WA 94.15 (egs2 reports 94.8).

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>