[#33252] DocDB: Reduce lock contention in LongOperationTracker

[craigsoules]

Summary

Tracks upstream issue yugabyte/yugabyte-db#33252 (to be upstreamed from this PR).

Every LongOperationTracker registration went through a single process-wide mutex in LongOperationTrackerHelper: Register locked the mutex, pushed into a shared std::priority_queue, and issued an unconditional notify_one (a futex wake on nearly every registration, since the checker thread is almost always waiting). The woken checker thread immediately re-acquired the same mutex, contending with concurrent registrations. This sits on hot paths: every read query, every Raft UpdateReplica, log append, master heartbeat processing, and (in debug builds) every ScopedRWOperation.

Make registration lock-free instead:

• Operations are handed to the checker thread through the lock-free MPSCQueue from yb/util/lockfree.h. Registration is now one allocation plus one CAS, with no mutex and no condvar notify.
• TrackedOperation ownership switches from std::shared_ptr to intrusive ref-counting (RefCountedThreadSafe + scoped_refptr) so the queue's raw pointer can carry an owned reference; the use_count() > 1 completion check becomes !HasOneRef().
• The checker thread drains the intake queue (bounded to 1000 per iteration so deadline processing and shutdown are never starved) into a private, uncontended priority queue and sleeps until the next deadline, capped at 100ms. Reporting is unchanged: each overdue operation still logs its own message, exact elapsed time, and stack trace.
• Deadlines shorter than 200ms (possible via runtime flags such as master_ts_heartbeat_long_operation_warning_ms) additionally wake the checker thread through a mutex-protected predicate, so short-deadline warnings cannot be lost; typical (1s+) registrations never touch the mutex.
• Fix stop_ being uninitialized.

Registration throughput on macOS arm64 (8 threads x 200k registrations): 4.2M/sec before vs 3.7-4.4M/sec after, i.e. parity within run-to-run noise on that platform. The expected production win is eliminating the per-registration futex wake and the mutex convoy between registering threads and the checker thread on Linux; a Linux perf/futex comparison is still to be collected.

Test plan

• ./yb_build.sh release --cxx-test debug-util-test --gtest_filter DebugUtilTest.LongOperationTracker (also debug)
• ./yb_build.sh release --cxx-test debug-util-test --gtest_filter DebugUtilTest.LongOperationTrackerStress (new: concurrent registration/move churn; also debug)
• ./yb_build.sh release --cxx-test debug-util-test --gtest_filter DebugUtilTest.LongOperationTrackerUnderLoad (new: overdue operations detected under sustained registration traffic and a 50k burst exercising the bounded drain; exact warning counts validate move construction/assignment clearing; also debug, and 10x repeated in release)
• ./yb_build.sh release daemons (all call sites compile)
• Linux sanitizer (ASAN/TSAN) runs via CI (Register is a no-op under sanitizers, pre-existing; macOS prebuilt thirdparty lacks instrumented libc++ locally)