fix(sarvam): align Bulbul validation with API constraints

[mikemikimike]

Fixes #6774

Summary

• Align the Bulbul v3 speaker compatibility list with the documented API speakers.
• Validate pace against the selected model (v2: 0.3-3.0, v3/v3-beta: 0.5-2.0), including model switches through update_options().
• Reject REST-only sample rates before opening a streaming session while keeping them valid for REST synthesis.
• Add provider-free regression coverage for all of the above.

Tests

• .venv\\Scripts\\python.exe -m pytest -q --disable-warnings livekit-plugins/livekit-plugins-sarvam/tests/test_tts_validation.py (33 passed)
• .venv\\Scripts\\python.exe -m pytest -q --disable-warnings livekit-plugins/livekit-plugins-sarvam/tests/test_speech_timing.py (12 passed)
• ruff format --check (passed)
• ruff check (passed)
• mypy livekit-plugins/livekit-plugins-sarvam/livekit/plugins/sarvam/tts.py (passed)
• git diff --check (passed)

The existing test_language_probability.py baseline currently fails before this change because its handcrafted SpeechStream fixture does not initialize _pending_eos (16 failures); it is unrelated to this TTS change.

[CLAassistant]

All committers have signed the CLA.