fix(tracing): preserve subagent trace context

[linhongyu510]

Summary

Fixes #549.

The in-memory span processor attached custom invocation and agent values to the OpenTelemetry context, but did not set the started span as the current span. A nested agent could therefore start without the parent agent span and receive an independent trace ID.

This change:

• places invocation and agent spans into the attached OTel context with set_span_in_context
• stores context tokens by span ID because on_end receives a ReadableSpan
• detaches the matching token when each span ends, restoring the parent context
• cleans up tokens for RECORD_ONLY spans even though they are not exported
• keeps cleanup in finally so exporter failures cannot leak context

Validation

tests/tracing/test_inmemory_exporter.py now covers:

• invocation → parent agent → child agent trace inheritance
• concurrent asyncio task isolation
• exceptional agent termination
• nested non-agent spans
• ReadableSpan cleanup in on_end
• unsampled RECORD_ONLY cleanup

Results:

• target tests — 6 passed
• Ruff check and format check
• git diff --check

AI assistance

This change was developed with AI assistance. I reviewed the ContextVar lifecycle, verified the patch with real OpenTelemetry spans, and ran the checks above.