feat(sync): push-style live sessions via open_live()

[ccampbell-aai]

Stacked on #241, which now carries the transcribe_live() rename and README examples. This PR adds only the push-style session on top.

Why

transcribe_live() takes an iterator, but most live audio arrives in a callback: sounddevice, PyAudio, WebRTC tracks, telephony media streams, websocket frame handlers. Every such caller had to build the same bridge before getting a transcript: a queue, a stop flag, a drain generator and, on the sync side, a thread to run the blocking call in. That bridge is the same thirty-odd lines in every application, and none of it is the application's job.

config = aai.SyncTranscriptionConfig(sample_rate=16000, channels=1)

with aai.SyncTranscriber() as transcriber:
    with transcriber.open_live(config) as session:
        with sd.RawInputStream(samplerate=16000, channels=1, dtype="int16",
                               callback=lambda data, *_: session.write(bytes(data))):
            input("Recording, press Enter to stop... ")
    print(session.result().text)

What

• SyncTranscriber.open_live(config) -> LiveSession; AsyncSyncTranscriber.open_live(config) -> AsyncLiveSession. Both exported from assemblyai.sync.v1.
• The request starts the moment the session opens (on the transcriber's existing worker pool, or as a task on the running loop), so the connection and config part go out while the speaker is still talking.
• write(chunk) queues audio and never blocks. Thread-safe on the sync side; on the async side it is a plain function so a callback can call it, and must run on the loop thread (loop.call_soon_threadsafe(session.write, chunk) from elsewhere). Chunk types are validated at write() so a str surfaces immediately rather than in result().
• close() ends the audio; result() waits for the transcript (closing first if needed); abort() drops the request. Context-manager exit closes on success and aborts if the block raised.
• Both are thin layers over transcribe_live(): no new transport code, same errors, same 120 s cap and silence-abort semantics. The iterator form stays for sources that already are iterators or file objects.

Abort works by having the producer raise a private exception, which propagates out of httpx and drops the connection; abort() swallows it. After close() the upload is already complete, so an abort then waits out the in-flight response (documented).

README now leads the live examples with the session (sounddevice, and a websocket handler for asyncio) and keeps transcribe_live() for iterator and file-object sources.

Validation

• 528 unit tests pass (19 new); the sync modules also pass on the httpx 0.22 floor. Ruff and mypy clean.
• Verified live against prod: sync and async sessions fed from a sounddevice capture callback return transcripts identical to transcribe() on the same PCM, with the post-speech wait cut to roughly a third; a mid-upload abort() returns immediately and drops the request.

No version bump: this stacks on #241 and inherits its version, so it carries the skip-version-bump label. The test matrix and version-bump workflows only trigger for PRs into master, so only lint runs here until the base moves.

🤖 Generated with Claude Code