performance: Implement arena allocator for hashtable (#2)

[AlexJuca]

Context

The current hash table allocates entries and values individually. Under high write volume this can create allocator overhead, fragmentation, and poor cache locality. An arena allocator could improve throughput and memory behavior, but it must preserve correctness for delete/update/free paths.

Scope

• Design an arena or pool allocator for hash table entries, keys, and/or value metadata.
• Define ownership for value payloads and deletion semantics.
• Support safe updates and deletes without use-after-free.
• Measure memory fragmentation and throughput before/after.
• Ensure sanitizer and leak tests remain clean.

Acceptance criteria

• Benchmarks show a measurable win or document why the arena approach is not worth merging.
• Delete/update/free paths are covered by tests.
• Memory accounting remains compatible with Add memory limits and eviction policy #73.
• Implementation does not make future persistence or eviction work harder.