Add E2E wall-clock timing for store requests (GPU IPC + async data paths)

[Copilot]

Adds a unified end-to-end timing mechanism in LMCacheMPWorkerAdapter that measures true wall-clock time from submit_store_request (forward thread) to when get_finished observes the future resolved — capturing actual GPU copy completion, not just kernel launch time.

Changes

• import time added to standard imports
• _store_submit_times: dict[str, float] added to LMCacheMPWorkerAdapter.__init__ to track per-request submit timestamps
• submit_store_request: records time.perf_counter() at entry, keyed by request_id
• get_finished: when s_future.query() returns True, pops the submit timestamp and emits:

  [E2E-STORE] req=<id> e2e=X.XXX ms
  

• Unhealthy drain path: _store_submit_times.clear() alongside the existing store_futures.clear() calls

Why this works for both paths

Path	What the future resolves after
GPU IPC (HandleTransferContext)	MQ latency + server kernel + D2H copy + event sync + poll
Async data (AsyncDataTransferContext)	Thread scheduling + gather launch + actual GPU copy (sync) + commit + poll

Both paths surface a MessagingFuture polled in get_finished, so a single instrumentation point gives an apples-to-apples comparison. Grep [E2E-STORE] to extract timings.