fix(voice): keep trace active until pipeline processing completes

[weiguangli-io]

Summary

• move TraceCtxManager scope into the background tasks for both single-turn and multi-turn voice pipeline execution so the trace stays open for the full async lifecycle
• avoid canceling the producer task immediately after session_ended; wait for graceful producer completion first so trace_end can be emitted reliably
• add voice pipeline regression tests for:
  • trace not finishing early in single-turn mode
  • trace ending after multi-turn processing
  • on_start() intro/exception path continuing without aborting the session

Root Cause

VoicePipeline previously entered TraceCtxManager only around task creation, then returned immediately. This ended the trace before async processing finished, causing subsequent spans to be emitted after trace closure.

Testing

• uv run ruff check src/agents/voice/pipeline.py src/agents/voice/result.py tests/voice/test_pipeline.py
• env -u ALL_PROXY -u all_proxy -u HTTPS_PROXY -u https_proxy -u HTTP_PROXY -u http_proxy -u NO_PROXY -u no_proxy uv run --extra voice pytest tests/voice/test_pipeline.py -q
• env -u ALL_PROXY -u all_proxy -u HTTPS_PROXY -u https_proxy -u HTTP_PROXY -u http_proxy -u NO_PROXY -u no_proxy uv run --extra voice coverage run -m pytest tests/voice/test_pipeline.py -q
• changed-line coverage check: src/agents/voice/pipeline.py and src/agents/voice/result.py are both 100%

Fixes #2470