fix: use parametrizations.weight_norm to silence deprecation warning - #3087

Fixes #3074

What & why

torch.nn.utils.weight_norm is deprecated in favor of
torch.nn.utils.parametrizations.weight_norm, which uses the modern
parametrization API. The old API emits a FutureWarning at import/use time
and is scheduled for removal.

SpeechBrain called the deprecated API directly in three source files:

• speechbrain/nnet/CNN.py — Conv1d / Conv2d / ConvTranspose1d
• speechbrain/lobes/models/HifiGAN.py — DiscriminatorS
• speechbrain/lobes/models/beats.py — positional conv

speechbrain/lobes/models/discrete/dac.py already uses the new API via its
try/except import, so it needs no change.

Fix

Replace nn.utils.weight_norm(...) with
nn.utils.parametrizations.weight_norm(...) in the three files above. The new
API is documented as state_dict-compatible with the old one, so no checkpoint
migration is required. remove_weight_norm() calls are left untouched — in
current PyTorch it transparently handles parametrized modules.

Verification

• No nn.utils.weight_norm( remains in speechbrain/ source (excl. recipes/tests).
• The change is a pure 1:1 API swap on the same module, same signature
  (name/dim args in beats.py are preserved).
• I could not run the full test suite locally (PyTorch unavailable in this
  environment), so behaviour was not exercised end-to-end here; CI is expected
  to validate forward/backward and checkpoint loading.

Note

This is a deprecation-warning fix rather than a crash — it does not alter
current runtime behaviour, but prevents breakage once the old API is removed.