feat(prerecorded): add AsyncTranscriber for asyncio callers - #227

Closes #80.

`Transcriber.transcribe_async` returns a `concurrent.futures.Future`. That
future is not awaitable, and `.result()` blocks the event loop. An asyncio
caller must therefore route transcription through a thread pool, which the
hardware bounds.

`AsyncTranscriber` provides the same API as `Transcriber`, with coroutines. It
takes the same `TranscriptionConfig`. Its result has the same fields. It needs
one thread.

The new modules follow the layout of their threaded twins, one public class per
module, layered `async_transcript <- async_client`:

- `prerecorded/v2/async_transcript.py`: `AsyncTranscript`.
- `prerecorded/v2/async_client.py`: `AsyncTranscriber`.
- `prerecorded/v2/async_api.py`: an asyncio version of each function in
  `api.py`, raising the same error per endpoint through `_base`.
- `async_client.py` at root holds `AsyncClient`, which wraps an
  `httpx.AsyncClient`. It sits beside `Client`, because `prerecorded/v2` and
  `sync/v1` both use the root client. There is no process-wide default: the
  pool belongs to the event loop that first used it, so a global pool fails on
  a second `asyncio.run()`.

Behavior notes:

- `AsyncTranscript` has the same fields as `Transcript`. Only the methods that
  call the API are coroutines. The transcriber provides `get_by_id` and
  `delete_by_id`, because the async transcript needs the transcriber's pool.
- An upload sends a path or a file object in chunks. A worker thread reads each
  chunk. The request sets `Content-Length` when the size is known.
- `transcribe_group` and `submit_group` keep the input order, limit concurrent
  work to `max_concurrency`, and report every error. The threaded group methods
  discard errors when `return_failures` is not set.
- LeMUR remains synchronous only. `LemurSource` now accepts an
  `AsyncTranscript`.

The tests avoid every pytest-httpx API added after 0.20, because the pinned
httpx tox envs resolve pytest-httpx back that far. The group tests stub
`async_api.create_transcript` rather than the HTTP layer, since no pytest-httpx
version in the matrix can delay a mocked response.

Tests: 45 tests in `tests/unit/test_async_transcriber.py`. The full suite gives
433 passed, 3 failed on httpx 0.22, 0.24, and 0.28, and under both pydantic
paths. The 3 failures need `pyaudio` and also fail on the base branch.

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