feat: add configurable cache statistics and latency snapshots

[leiysky]

Summary

Add RuntimeOptions::stats and Cache::stats_snapshot() so applications can collect complete request outcomes and cumulative latency distributions independently of the legacy statistics switch. L1 hits, L2 lookups and mutations each support disabled, full or sampled timing. L2 timing starts at L1 miss, allowing full L2 collection without clocks on L1 hits; engine timing covers read/write/reclaim requests from slot reservation through completion.

Recorders use bounded atomic stripes charged to managed memory. Structured snapshots expose timing scope, sampling mode, bucket counts, duration sums and validity; applications own metric conversion and export. Fully specified RuntimeOptions literals need a stats field or a struct-update default. Includes a runnable snapshot example, benchmark controls and regression coverage for accounting, timing boundaries, cancellation and concurrency.

Validation: cargo x check, cargo x test, and cargo x lint passed. The io_uring library also cross-compiles for x86_64-unknown-linux-gnu; Linux runtime execution was not tested locally.