Replace per-step store timing logs with unified StoreTimer utility

[Copilot]

The four store paths (GPU IPC C++ ops, GPU IPC Python ops, Async SHM, Async Pickle) each emitted verbose per-step logger.info timing lines ([FWD ...], [BG ...], [GPU-STORE], [PICKLE-COMMIT], [SHM-PREPARE], etc.) unconditionally at INFO level, adding noise and time.perf_counter() overhead on every request in production.

Changes

New: lmcache/v1/multiprocess/store_timer.py

StoreTimer — zero-overhead timing utility gated on logger.isEnabledFor(logging.DEBUG) at construction. No timestamps are taken in production. Emits a single [STORE-TIMING] debug line with four standardized metrics across all paths:

[STORE-TIMING] req=<id> path=<gpu_ipc|shm|pickle>
    fwd_return=X copy_submitted=X kv_releasable=X e2e_complete=X ms

Metrics:

• fwd_return — time until the vLLM forward thread is unblocked
• copy_submitted — time until all GPU copies are enqueued and CPU returns
• kv_releasable — time until GPU has finished reading paged KV (safe to evict)
• e2e_complete — time until data is committed and retrievable

Integration

• gpu_transfer.py — path="gpu_ipc": marks fwd_return at entry (FWD already returned on server side), copy_submitted after event.record(), kv_releasable at gpu_sync point, e2e_complete after submit_callback_to_stream for finish_write
• async_data.py — path="shm" or "pickle": marks fwd_return when submit_store returns to the forward thread; background closure marks copy_submitted after stream exits, kv_releasable after gather_done.synchronize(), e2e_complete after commit_store()

Cleanup

• Removes all verbose per-step logger.info timing blocks from gpu_transfer.py, async_data.py, non_gpu_transfer.py, and server_transfer.py
• Removes now-unused time.perf_counter() variables; removes import time from server_transfer.py
• Retains existing "Stored %d tokens in %.3f seconds" info-level logs in gpu_transfer.py and non_gpu_transfer.py