perf: huge pages (MADV_HUGEPAGE) for the store table + arenas

[ELares]

Parent: #507. Cheap (est. 3-10%, low risk); do early.

Problem

The -r 1M random-key benchmark (and real skewed workloads) thrash the TLB: each GET touches a random table bucket + a random value, mostly TLB misses with 4 KiB pages. 2 MB huge pages cut the TLB-miss rate dramatically for the same coverage.

Mechanism / fix

madvise(MADV_HUGEPAGE) (transparent huge pages) on the per-shard table backing store + the value arenas, or allocate them from a huge-page-backed arena. Gate behind Linux + a config flag (default on where THP is available; a no-op elsewhere). This is a memory hint, so it does not touch ADR-0003 determinism.

Risk / notes

• THP can add memory overhead (2 MB granularity); measure RSS and make it configurable.
• Interacts with the memory-determinism moat goal ([EPIC] Close the Dragonfly GET gap + own the tail-latency/determinism niche #507): huge pages also help RSS predictability.

Acceptance

• Table/arena backed by huge pages on Linux (verify via /proc/PID/smaps AnonHugePages).
• No functional change; tests pass.
• A/B target: 3-10% GET on the random-key workload.