Fix StoreTimer start point and async path labeling

[Copilot]

This adjusts multiprocess store timing so submit_store() measures from the caller-visible entry point, and removes an invalid async GPU timing mark that was being emitted without waiting for device completion.

• Start timing at submit_store() entry

  • AsyncDataTransferContext.submit_store() now creates StoreTimer immediately after the registration check, before prepare_store() and buffer allocation.
  • This makes _t0 reflect the latency seen by the caller, rather than starting partway through the store path.

• Support deferred path resolution

  • Added StoreTimer.set_path() so the timer can be created early and labeled later once the transfer mode is known.
  • async_data.py now starts with a generic path and updates it to shm or pickle after used_shm_direct is determined.

• Remove meaningless async GPU mark

  • Dropped timer.mark("kv_releasable") from the async GPU IPC store path in gpu_transfer.py.
  • That path does not synchronize on the completion event, so the mark was effectively stamped at submit time and conveyed no real device-completion signal.

• Example

  timer = StoreTimer(request_id, path="data")
  ...
  timer.set_path("shm" if used_shm_direct else "pickle")