Add request-scoped profiler benchmark flag and compressed trace export

[CrazyDave999]

Summary

(Based on sgl-project/mini-sglang#14, with some follow-up fixes.)

Clients can send "profile": true and mini-sglang will start a torch.profiler session for that request, then export a Chrome-trace artifact when the request finishes.

What changed

• API:
  Add benchmark-side support to pass profiling on demand (--profile), forwarding "profile": true in benchmark requests for easier validation.

• Messaging: propagate profile through tokenizer/backend messages (TokenizeMsg, UserMsg).

• Scheduler/core: carry profile into the internal Req object and start/stop profiling around the first in-flight request with profile=True.

• New utility: python/minisgl/utils/profiler.py implements RequestProfiler (start/stop/export).

Test

# Start server
export MINISGL_PROFILE_DIR=/tmp
python -m minisgl --model "Qwen/Qwen3-0.6B" --tp 2 --attention-backend fi

# Then on client side:
curl -X POST http://127.0.0.1:1919/generate \
  -H "Content-Type: application/json" \
  -d '{"prompt":"hello","max_tokens":64,"profile":true}'

# Or:
python benchmark/online/bench_qwen.py --profile

ls /tmp/minisgl*.json.gz