nlog: replace ConcurrentHashMap with phmap::parallel_flat_hash_map

[SBALAVIGNESH123]

Fixes #441

This PR replaces the current ConcurrentHashMap implementation in NativeLog, which uses a single std::shared_mutex around an absl::flat_hash_map, with phmap::parallel_flat_hash_map from the parallel-hashmap library. The old design meant every read and write operation competed for one global lock, which becomes a bottleneck under concurrent access.The new implementation uses phmap's internal submap-level locking — the map is split into 16 independent submaps, each with its own std::shared_mutex. This means operations on keys that hash to different submaps run fully in parallel with zero contention. Multiple readers on the same submap can also proceed concurrently thanks to the shared_mutex reader-writer separation.parallel-hashmap is a well-established header-only library built on top of Google's absl::flat_hash_map, widely used in production OLAP systems like Apache Doris. It was the library suggested in the original issue.All 15 public methods are preserved with identical signatures, so LogManager.cpp requires no changes. The only method removed is eraseIf, which was dead code — never called anywhere in the codebase. Template defaults (std::hash, std::equal_to) are kept the same to maintain backward compatibility with StreamIDShard's existing std::hash specialization.

Benchmark references from the original issue:

• https://martin.ankerl.com/2022/08/27/hashmap-bench-01/
• https://greg7mdp.github.io/parallel-hashmap/

Note: The .gitmodules merge conflict is due to new submodules added to develop since my fork was last synced. Happy to rebase onto latest develop if needed — just let me know. The actual code change (ConcurrentHashMap.h) merges cleanly.