feat(deepgram): make Flux audio chunk duration configurable

[dnaumenko]

Applications currently have to replace or patch SpeechStreamv2._run to evaluate Flux with an outgoing audio chunk size other than the hardcoded 50 ms. This adds a constructor-only chunk_size_ms option to deepgram.STTv2, retaining 50 ms as the default.

from livekit.plugins import deepgram

stt = deepgram.STTv2(
    model="flux-general-multi",
    chunk_size_ms=80,
)

Deepgram recommends 80 ms audio chunks for Flux. The setting controls outgoing mono PCM WebSocket batching.

The constructor rejects non-positive/non-integer durations and durations that yield fewer than one sample. The existing send loop still handles partial-chunk flushing and stream closure.

Validation

• uv run --no-sync pytest tests/test_plugin_deepgram_stt.py -q: 38 passed.
• New fake-WebSocket tests exercise the real public stream/send loop at 8/16/48 kHz, default 50 ms and explicit 20/80 ms chunks, byte-for-byte PCM preservation, partial tails, and closure ordering. Constructor validation is also covered.
• make fix and make check passed with UV_NO_SYNC=1: repository formatting/lint and mypy over 657 source files. The isolated environment includes the locked dev/typing groups plus optional boto3 and missing types-jsonschema/types-nanoid stubs.
• git diff --check passed. No live-provider performance test was run against this new patch.

Related proposal: #7204. Opening as a draft for API feedback.

[CLAassistant]

All committers have signed the CLA.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[rchengLK]

@dnaumenko could you resolve the merge conflicts first please?

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.