feat(amd): optionally process SIP early media

[swayamg20]

Summary

• add wait_until_answered to AMD, defaulting to True
• when disabled for a SIP participant, start detection timers and process audio and transcripts from track subscription
• preserve the existing answer-gated behavior by default
• cover the default, opt-in, and pending-answer cleanup paths

Why

Some carriers deliver voicemail greetings or announcements through 183 Session Progress early media before sip.callStatus becomes active. AMD currently discards that audio while waiting for the answered state.

This gives integrations an explicit opt-in without changing the safer default for calls where early media contains only ringback or dialtone. Callers that need a verdict before answer must also disable wait_until_answered on the SIP participant request.

Testing

• uv run pytest tests/test_amd_classifier.py -q — 37 passed
• uv run ruff check livekit-agents/livekit/agents/voice/amd/detector.py tests/test_amd_classifier.py
• uv run ruff format --check livekit-agents/livekit/agents/voice/amd/detector.py tests/test_amd_classifier.py
• scoped mypy for livekit.agents — 207 source files passed

Closes #6895

[devin-ai-integration]

✅ Devin Review: No Issues Found

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