[Profiler] Export torch profiler trace off the worker's main thread

[elvircrn]

Summary

Export trace on a background thread instead of inline in on_trace_ready, since Kineto has already stopped collecting by then. Verified against a real opt-125m server: stop_profile() 22.6s -> 3.5s, no correctness regressions.

(Earlier revision also made Executor.profile() dispatch non-blocking via collective_rpc(non_block=True). Dropped it — the returned Future only resolves if something else calls .result() on it, which nothing does for this call, so it hung under real multiprocessing. Worker-side fix alone already solves the reported slowdown.)

Repro

from vllm import LLM, SamplingParams
import time

llm = LLM(
    model="facebook/opt-125m",
    enforce_eager=True,
    gpu_memory_utilization=0.2,
    profiler_config={"profiler": "torch", "torch_profiler_dir": "/tmp/trace"},
)
llm.start_profile()
llm.generate(["The quick brown fox"], SamplingParams(max_tokens=32))

t0 = time.perf_counter()
llm.stop_profile()
print(f"stop_profile(): {time.perf_counter() - t0:.1f}s")

Measured on CPU, opt-125m, gzip'd trace:

• main: stop_profile() = 22.6s
• this PR: stop_profile() = 3.5s

Test plan

• New CPU-only tests in tests/v1/worker/test_gpu_profiler.py
• Verified against real facebook/opt-125m on CPU: start_profile -> generate -> stop_profile -> generate, no hang, real trace produced

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @elvircrn.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork