fix(mutex): hold lock during notify_all in Barrier::wait to prevent teardown race (fixes #1672)

[jdymitarai]

Root Cause

In Barrier::wait() (src/mutex.h), MutexLock ml(lock_) was enclosed in an inner block that released the mutex before calling phase_condition_.notify_all(). When other threads awakened by createBarrier() finish their work and immediately destroy or free the Barrier instance, the last thread proceeds to call phase_condition_.notify_all() on a deallocated or destructed condition variable, resulting in a segmentation fault / null pointer dereference (#1672).

Fix Approach

Extend MutexLock ml(lock_) to cover phase_condition_.notify_all(). Holding the lock until after notify_all() completes ensures the Barrier object and its condition variable remain valid and protected from premature destruction by waking peer threads.

Verification

Added regression tests to test/benchmark_gtest.cc:

• BarrierTest.SingleThreadedWaitAndTeardown: 100 iterations of single-threaded wait and immediate destruction.
• BarrierTest.MultiThreadedWaitAndTeardown: 100 iterations of 8-threaded barrier wait and immediate join/destruction.
• BarrierTest.MultiThreadedRemoveThreadAndTeardown: 100 iterations verifying removeThread() and concurrent teardown safety.
  All tests pass cleanly with zero sanitizer warnings.

[dmah42]

nice catch. thanks. can you fix the formatting?

[jdymitarai]

Updated the formatting in test/benchmark_gtest.cc to match clang-format. Thanks @dmah42!