feat(rollout): add rollout-side torch profiler trigger via sglang

[niu0421]

Add CLI-driven torch profiler support for the rollout (sglang) side, complementing the existing train-side --profile-target.

Currently, profiling sglang engines during rollout requires manually calling the /start_profile HTTP endpoint. This PR exposes the
existing SGLangEngine.start_profile() / stop_profile() API through RolloutManager, triggered at a user-specified rollout_id.

Changes:

• slime/ray/rollout.py — add _maybe_start_rollout_profile() / _stop_rollout_profile() in generate(), calling
  engine.start_profile.remote() / engine.stop_profile.remote() on all rollout engines
• slime/utils/arguments.py — add --rollout-profile-* arguments:
  • --rollout-profile-start-rollout-id: rollout id to trigger profiling (disabled when unset)
  • --rollout-profile-num-steps: number of sglang forward steps to profile
  • --rollout-profile-start-step: warmup steps to skip before recording
  • --rollout-profile-activities: CPU/GPU/MEM activities to capture
  • --rollout-profile-output-dir: trace output directory
  • --rollout-profile-by-stage: split prefill/decode into separate traces
  • --rollout-profile-with-stack: record Python stacks
  • --rollout-profile-record-shapes: record tensor shapes

Usage:
--rollout-profile-start-rollout-id 1
--rollout-profile-num-steps 10
--rollout-profile-output-dir /tmp/rollout_traces

Validation:
Tested with tests/test_qwen2.5_0.5B_async_short.py on 4×H20:

• Default disabled: no profile logs, training completes normally, no performance regression
• Enabled at rollout_id=1: profiler starts at correct rollout, each engine generates a trace.json.gz, profiler stops cleanly
• All engines log Start profile / Profiling done as expected