Add optional Silero VAD backend for longform transcription

[gman-dev-nov]

Motivation

transcribe_longform depends on the gated pyannote/segmentation-3.0 model, so the first run requires a Hugging Face account, accepting the model conditions, and an HF_TOKEN. That first step is a real hurdle in corporate and air-gapped environments where creating an HF account or reaching the Hub is not an option.

This PR adds an opt-in alternative: transcribe_longform(..., vad_backend="silero"), using Silero VAD, whose weights ship inside the silero-vad pip package — no account or token needed.

Relation to #24

#24 proposed replacing the pyannote pipeline and was declined. This PR is deliberately different:

• The default is unchanged — transcribe_longform(...) behaves exactly as before; the pyannote pipeline code in vad_utils.py is untouched.
• Each backend lives in its own module; the existing chunk-merging logic is shared by both instead of being reimplemented.

Changes

• gigaam/chunking_utils.py (new) — the span-merging logic moved verbatim out of segment_audio_file; it operates on plain (start, end) tuples, so it is independent of which VAD produced the spans. The four tuning knobs are grouped into a frozen ChunkingConfig dataclass.
• gigaam/vad_utils.py — pyannote code untouched; segment_audio_file keeps its public signature and delegates merging to chunking_utils.
• gigaam/silero_vad_utils.py (new) — Silero counterpart of segment_audio_file with a slimmer signature: (wav_file, sr, config: Optional[ChunkingConfig]) — no device argument, since Silero VAD always runs on CPU. Importable without pyannote installed; a missing silero-vad package raises an ImportError pointing to pip install gigaam[silero].
• gigaam/model.py — transcribe_longform gains a vad_backend: str = "pyannote" keyword (backwards-compatible) and dispatches to the backend module via match/case; the import was already function-local upstream. Unknown values are rejected with a ValueError before any audio is decoded.
• pyproject.toml — new silero extra (silero-vad==6.2.* — the tested minor, pinned in the repo's style; validated against 6.2.1); also added to tests so CI exercises the new path — happy to drop that if you prefer to keep the tests extra minimal (the new tests importorskip silero-vad, so they degrade to skips).
• tests/test_longform.py — 3 new tests that run without HF_TOKEN, giving longform coverage even where the secret is unavailable (e.g. PRs from forks). The e2e test compares merged text against the pyannote-based reference via similarity > 0.9 rather than exact boundaries, since chunk borders legitimately differ between VADs.
• README.md / README_ru.md — one sentence in the longform setup block.

Testing

• pytest tests/test_longform.py -k "silero or unknown_backend" — 3 passed (Apple M-series: MPS and CPU; silero-vad 6.2.1).
• black --check, isort --check-only, flake8 (workflow flags), mypy --ignore-missing-imports --no-strict-optional — all clean.
• Verified in a pyannote-free environment: gigaam.silero_vad_utils imports, vad_backend="silero" transcribes end-to-end.
• Real-world check: the 71-second sample transcribed with vad_backend="silero" on MPS in ~6 s, merged text matching the pyannote-based reference.

Note on CI: the pre-existing pyannote longform tests will be red on this PR — GitHub does not expose secrets.HF_TOKEN to pull requests from forks. The new silero tests are unaffected and should pass.

[gman-dev-nov]

@Bobrosoft98 @georgygospodinov @sverdoot @nshmyrev watch please 🙏