perf(async_data): move gather phase off forward thread into background executor

[Copilot]

submit_store was launching gather_paged_kv_to_cpu on the forward thread while the copy stream had a pending event-wait for the forward pass. CUDA runtime back-pressures the CPU when queuing kernels to a stream with unresolved dependencies, blocking the forward thread ~38ms per store.

Changes

• Move gather entirely to background thread: The with torch_dev.stream(self._copy_stream): block — _event.wait(), gather_paged_kv_to_cpu(), and gather_done.record() — is now inside _commit_after_gather running on commit_executor. submit_store only does prepare_store + buffer allocation, then submits and returns immediately.

• _inflight_gather_events tracking preserved: gather_done is added to the set in the background thread under lock, after record(), before synchronize(). flush_inflight_gathers() semantics are unchanged.

• Remove profiling remnants: Dropped import time, all t00/t1/.../t11 timing variables, the two torch_dev.synchronize() calls, and all [Store Profiler] log lines that were left in from profiling sessions.

Before (forward thread blocked):

with torch_dev.stream(self._copy_stream):
    _event.wait(stream=self._copy_stream)
    gather_paged_kv_to_cpu(...)   # ~38ms CPU wall time on forward thread
    gather_done = torch_dev.Event()
    gather_done.record(self._copy_stream)
# submit commit to background
commit_executor.submit(_commit_after_gather)

After (forward thread returns immediately):

def _commit_after_gather() -> None:
    with torch_dev.stream(self._copy_stream):
        _event.wait(stream=self._copy_stream)
        gather_paged_kv_to_cpu(...)   # runs fully off forward thread
        gather_done = torch_dev.Event()
        gather_done.record(self._copy_stream)
    with self._inflight_lock:
        self._inflight_gather_events.add(gather_done)
    gather_done.synchronize()
    non_gpu_context.commit_store(...)

commit_executor.submit(_commit_after_gather)  # submit_store returns here