feat: add per-step timing instrumentation for non-GPU server store path

[Copilot]

The non-GPU (async data) server-side store path only emitted a single "Stored N tokens in X seconds" log with no per-step breakdown, unlike the GPU IPC path which has detailed stage timings.

New log output

non_gpu_transfer.py — outer module layer:

[SRV-PREPARE-STORE] req=<id> resolve_keys=0.012 prepare=1.234 total=1.246 ms (strategy=shm)
[SRV-COMMIT-STORE]  req=<id> commit=0.987 total_since_prepare=2.345 ms (strategy=shm, num_tokens=256)

server_transfer.py — pickle path:

[PICKLE-COMMIT] req=<id> deserialize=0.543 reserve_write=0.211 copy_loop=1.102 finish_write=0.089 total=1.945 ms (num_chunks=4)

server_transfer.py — SHM path:

[SHM-PREPARE] req=<id> resolve_keys=0.234 reserve_write=0.156 slots=0.043 total=0.433 ms (num_slots=4)
[SHM-COMMIT]  req=<id> finish_write=0.078 total=0.134 ms (num_keys=4)

Changes

• TransferStrategy (ABC): adds abstract strategy_name property; PickleTransferStrategy returns "pickle", ShmTransferStrategy returns "shm".
• PickleTransferStrategy.commit_store(): instruments pickle.loads → reserve_write → tensor copy loop → finish_write with [PICKLE-COMMIT].
• ShmTransferStrategy.prepare_store(): instruments resolve_obj_keys → reserve_write → slot-descriptor loop with [SHM-PREPARE].
• ShmTransferStrategy.commit_store(): instruments finish_write on the SHM fast path (non-fallback branch) with [SHM-COMMIT]; fallback goes through [PICKLE-COMMIT].
• NonGPUTransferModule.prepare_store(): wraps strategy.prepare_store() with [SRV-PREPARE-STORE] (context-lookup + strategy-call breakdown).
• NonGPUTransferModule.commit_store(): wraps strategy.commit_store() with [SRV-COMMIT-STORE] including total_since_prepare end-to-end latency.

All timings use time.perf_counter(); all log calls use %s/%.3f format strings. No functional logic changed.