feat: add direct async iteration on Dataset and KeyValueStore - #2267

Add a non-abstract `KeyValueStoreClient.iterate_entries` to the storage
client base class, yielding `KeyValueStoreRecord`s. The default
implementation lists the keys with `iterate_keys` and reads each value
with `get_value` as the iteration advances, so existing and third-party
clients need no changes. Backends that can read keys together with their
values more efficiently can override it.

`KeyValueStore.iterate_entries` and `iterate_values` now delegate to the
client method instead of combining `iterate_keys` and `get_value`
themselves.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wg6jZgyQp7XkvVvxuueJyV

Override `iterate_entries` where the backend can do better than one
value read per key:

- The SQL client selects keys, metadata and values in a single streamed
  query, so a store with N records costs one query instead of N+1 and
  still holds only one row in memory at a time.
- The Redis client fetches values with one HMGET per batch instead of
  two round trips per record. Batches are bounded by key count and by
  the record sizes known from the metadata hash, so large values do not
  pile up in memory. A record larger than the byte limit is fetched
  alone.

Both clients share the value decoding with `get_value` through a new
`_build_record` helper.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wg6jZgyQp7XkvVvxuueJyV

…e client

`set_value` stores an empty byte string for `None`, so the batched read
can fetch those records like any other and let `_build_record` map the
`application/x-none` content type to `None`. This drops the key filtering
and the per-key lookup dict from the batch fetch.

Also reword the base `iterate_entries` docstring so it does not present
the handling of records deleted mid-iteration as a contract.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wg6jZgyQp7XkvVvxuueJyV

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wg6jZgyQp7XkvVvxuueJyV

…ype checker

redis-py types every reply as `bytes | str | None` because a client
created with `decode_responses=True` returns strings. The key-value
store client stores binary values and needs the raw bytes back, so it
narrowed the type with `ty: ignore` and would fail with an
`AttributeError` on such a client.

Add `expect_bytes`, which narrows a reply to bytes and raises a
`TypeError` naming the unsupported option, use it at both value-reading
sites, and document the requirement on `RedisStorageClient`.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wg6jZgyQp7XkvVvxuueJyV

`expect_bytes` now handles a single reply, which `isinstance` narrows
on its own, and the HMGET result is narrowed with a comprehension.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wg6jZgyQp7XkvVvxuueJyV

…ient

Datasets use RedisJSON and request queues parse JSON strings, so both
work with a decoding Redis client. Only key-value store values are
binary.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wg6jZgyQp7XkvVvxuueJyV

`stream_results` buffers up to 1000 rows client side, so iterating a
store of large values held hundreds of megabytes before the first yield,
and the single open transaction lasted for the whole iteration.

The SQL client now reads record metadata in keyset-paginated pages,
splits each page into batches bounded by the record sizes, and fetches
every batch's values with one query. Each query runs in its own short
session. The batching rule is shared with the Redis client through a new
`batch_records_by_size` helper.

Measured with 300 records of 1 MiB: peak memory drops from 300 MiB to
9 MiB. With 2000 small records the iteration issues 41 selects instead
of 2001 and is about 40 times faster than the default implementation.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wg6jZgyQp7XkvVvxuueJyV

The test directories have no `__init__.py`, so pytest cannot import two
modules named `test_utils.py` and aborted the whole unit test collection.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wg6jZgyQp7XkvVvxuueJyV