refactor: reduce loading overhead tracker lock contention

[sunby]

Summary

This PR reduces lock contention in LoadingOverheadTracker by splitting the previous single global mutex into:

• a std::shared_mutex for group/handle map access
• a per-group GroupState mutex for reservation state updates

The hot-path operations (Reserve, Release, HasFiniteUpperBound, and GetUpperBound) now take a shared map lock, copy the target group state pointer, and only serialize updates within that group. This allows independent groups to reserve/release loading overhead without blocking each other on one global tracker mutex.

Registration and cleanup behavior is preserved:

• existing groups still reuse the same handle and increment ref counts
• finite upper bounds still take the max per dimension when re-registered
• unlimited upper bounds can still be replaced by the registered bound
• Unregister still removes the group at ref count 0 and logs residual reservations before cleanup

Testing

• pre-commit passed
• UT on ubuntu-22.04 passed
• UT on macos-15 passed