perf(worker): profile and reduce cache-hit logging, allocation, and lock overhead

[somfornot]

Problem / motivation

Once bulk bytes use sendfile, protocol scheduling and shared userspace bookkeeping determine the next ceiling. Several operations occur around every resident hit and may constrain multi-ring scaling:

• Whole/paged sendfile hits emit INFO logs (runtime.rs); worker and load-test defaults are RUST_LOG=info (main.rs, dataplane_loadtest.sh). At 100K+ requests/s this can dominate user work or contaminate benchmarks.
• BlockIndex has one process-wide RwLock (index.rs).
• LRU touch/pin/unpin uses one process-wide Mutex; eviction scans tracked units while holding it (eviction.rs).
• The version cache uses one Mutex; headers, frame rejoin, and some byte paths allocate/copy per request.

These are hypotheses, not a reason to add sharding or lock-free complexity without profiles. Talon's previous request-affinity experiment regressed performance.

Proposed work

1. Profile all-hit workloads with logging disabled, default logging, and realistic sinks.
2. Move per-hit success events to DEBUG/TRACE or sampling while retaining counters and slow/error logs.
3. Measure lock wait/hold time, allocations, cycles/request, cache misses, context switches, and ring scaling.
4. Optimize only measured bottlenecks. Candidates include deferred/approximate LRU touches, shorter eviction critical sections, sharded metadata, reusable frame buffers, and removal of payload-sized copies.
5. Preserve fd pinning, eviction, version freshness, recovery, and cardinality bounds.

Acceptance criteria

• A reproducible profile attributes CPU/allocation/lock cost to logging, BlockIndex, LRU, version cache, framing, and metrics.
• Default INFO logging no longer emits one event per successful cache hit.
• Benchmark scripts set and record logging configuration explicitly.
• Any data-structure rewrite has contention tests and preserves eviction/pinning/recovery semantics.
• Before/after results cover 1/physical/logical ring counts and report throughput, user/system CPU, allocations, contention, and p99/p999.
• Unmeasurable or negative changes do not become default.

Related: #285, #291, #303, #569