[Feature] Instrument sandbox agents and SessionServer tracing

[matrix72c]

Summary

Add OpenTelemetry instrumentation to the rollout, sandbox, and SessionServer paths so a request trace remains connected across Ray actors, sandbox stages, and the proxy lifecycle.

Motivation

The existing rollout and SessionServer paths already expose the live boundaries needed for tracing, but their OTel context was not consistently propagated across Ray calls or represented in sandbox and proxy spans. This made cross-process request traces incomplete even though the underlying training behavior was correct.

Implementation

• Propagate the active OTel context through Ray generate_group calls for a single RolloutState or a concrete list/tuple group, while restoring internal carrier and call-chain fields after the call.
• Bridge the existing sandbox stage timer to live OTel spans with stable names and a bounded attribute schema. Legacy JSONL enter/exit records and timing semantics remain unchanged.
• Trace the SessionServer proxy lifecycle as prepare_request, backend_roundtrip, and record_response under one request span. Incoming aiohttp headers are used as the parent carrier, and the backend carrier is injected inside the roundtrip span.

Compatibility

• No training behavior or request/response payload format changes.
• No prompt, response body, or raw header attributes are exported.
• No duplicate duration/ok telemetry is added; OTel owns duration and logical failures use error=true.
• This PR instruments existing live call sites only. It does not add synthetic trace reconstruction, trajectory polling, timeline storage, or a flush API.

Tests

• Focused rollout propagation, nested cleanup, sandbox bridge, SessionServer, and mixed-case W3C carrier tests.
• Existing SingleTurnAgentLoop and producer compatibility tests: 29 passed.
• ruff check on all changed Python files.
• compileall on changed modules and tests.
• git diff --check and final diff review against upstream/main.