Add Langfuse tracing for voice sessions

[Johnkothapalli]

Summary

• add optional Langfuse v4 tracing to the QuickVoice LiveKit voice runtime
• capture voice-session metadata and privacy-aware user/agent transcript observations
• record completion and turn-count evaluator observations plus native Langfuse scores
• add a smoke script, setup documentation, and a focused recording guide

Privacy and reliability

• transcript content is redacted by default
• zero-PII retention always redacts content while preserving non-sensitive turn counts
• Langfuse failures do not interrupt voice sessions
• failed session starts finalize with an error evaluation instead of leaving an open trace

Verification

• python -m py_compile apps\\ai\\handlers\\langfuse_handler.py apps\\ai\\handlers\\transcript_collector.py apps\\ai\\main.py apps\\ai\\scripts\\langfuse_trace_smoke.py apps\\ai\\tests\\test_langfuse_handler.py apps\\ai\\tests\\test_transcript_collector.py
• python -m unittest apps.ai.tests.test_langfuse_handler apps.ai.tests.test_transcript_collector (9 tests passed)
• live smoke trace ingested by Langfuse: langfuse-smoke-1784888287
• API read-back confirmed call_completed = 1.0 (BOOLEAN) and transcript_turn_count = 2.0 (NUMERIC)

The smoke script exercises the same LangfuseCallTracer used by apps/ai/main.py; a full phone call is not required to verify Langfuse ingestion.