perf: add thread-local TraceCache to skip redundant TraceTree lookups

[azurezene]

In allocation-heavy workloads, every malloc/free must walk the shared TraceTree trie to resolve its backtrace. This lookup holds a lock and consumes significant critical-section time, causing contention under multi-threaded workloads.

However, high-frequency allocation paths tend to be highly repetitive — the same call site allocating in a loop, the same function calling malloc repeatedly. The backtrace is identical every time, yet we pay the full cost of walking the trie on each call.

This PR inserts a thread-local 8-entry LRU cache in front of TraceTree::index() to skip redundant trie traversals. A cache hit avoids both the trie walk and the associated lock contention entirely.

Implementation:

Key: hash of the backtrace (Murmur-style via 0xbb67ae8584caa73b), with memcmp verification of the first 4 IP frames to eliminate false positives from hash collisions.
Eviction: decay-based rather than exact LRU — each miss decays the tail entry's hit counter by 25%; entries drop below the replacement threshold after ~4 consecutive misses. Hot entries resist eviction naturally.
Cache-line aware layout: key[] in the first cache line (read on every lookup), hitCount/index in the second (touched only on hit), IP storage in colder lines — minimizing cache misses on the fast path.
Lock-free: thread_local storage means zero contention, unlike any shared cache structure.

Performance：

This is the flame graph proportion for /heaptrack/tests/manual/threaded.cpp. It can be clearly seen that the proportion of the TraceTree::index function within the heaptrack_malloc function has decreased from 31.1% to 0.4%.

without cache:

with cache:

I found This idea was originally suggested in #13 but never landed