perf(alloc): use jemalloc to reduce memory fragmentation

[EddieHouston]

Summary

• Replaces the system allocator (glibc malloc) with jemalloc via tikv-jemallocator
• Reduces RSS by ~18% by eliminating malloc fragmentation that accumulates over time with RocksDB's allocation patterns
• Particularly important in memory-constrained environments where fragmentation inflates resident memory toward OOM

Background

glibc malloc fragments heavily under RocksDB's allocation pattern of many small, long-lived objects (block cache entries, memtable keys). Over hours of operation, RSS grows well beyond the logical heap size. jemalloc's arena-based allocator
and thread-local caching avoid this fragmentation.

Measured on a production-sized mainnet node (62 GB RAM, 1.45 TB DB):

• glibc malloc: RSS steadily climbed over time
• jemalloc: RSS stable at ~18% lower than glibc baseline

Test plan

• cargo check / cargo build with default features
• cargo build --features liquid
• Verify jemalloc is active (RSS should be noticeably lower than glibc on a mainnet node)
• No functional changes — allocator swap only