bench(engine): profile ReorderBuffer cache behavior at 1M items (#1854)

[oferchen]

Summary

• Adds crates/engine/benches/reorder_buffer_cache.rs, a Criterion bench that exercises ReorderBuffer insert/drain at 1M out-of-order indexed items.
• Varies item payload size {32 B, 256 B, 4 KB} and insertion order pattern {fully reverse, random shuffle, near-in-order with 10% deltas}; reports ops/sec via Throughput::Elements.
• Top-of-file documents how to run with perf stat -e cache-misses,cache-references (Linux) and valgrind --tool=cachegrind (any platform), plus what favorable / unfavorable counter results would imply for the storage layout (keep Box<[Option<T>]> vs. switch to flat Vec<T> + occupancy bitmap, or hot/cold split).

Notes

• Bench is gated behind BENCH_REORDER_CACHE=1 so default cargo bench runs remain fast - per-cell wall time is 5-30 s at 1M items.
• Payloads are pre-built outside the timed window via iter_batched(PerIteration) so the measurement reflects ReorderBuffer operations, not allocator work.
• Capacity is sized per pattern so no insert ever triggers grow(); the steady-state ring is what gets profiled.
• No changes to ReorderBuffer production code.

Test plan

• CI: fmt + clippy + nextest pass on Linux, macOS, Windows.
• Operator: BENCH_REORDER_CACHE=1 cargo bench -p engine --bench reorder_buffer_cache completes on a Linux host.
• Operator: perf stat -e cache-misses,cache-references,L1-dcache-load-misses,LLC-load-misses numbers captured for the 1M cells (driven from instructions in the bench file header).
• Operator: cachegrind run on at least one cell (e.g., payload256/shuffle) annotated via cg_annotate.