[TSL] Wait for queued and running work in default UnboundedWorkQueue::~UnboundedWorkQueue

[copybara-service]

[TSL] Wait for queued and running work in default UnboundedWorkQueue::~UnboundedWorkQueue

In the default/OSS implementation of tsl::UnboundedWorkQueue (xla/tsl/platform/default/unbounded_work_queue.{h,cc}), ~UnboundedWorkQueue() previously set cancelled_ = true immediately without waiting for pending or in-flight work items to finish, and then acquired thread_pool_mu_ while joining thread_pool_ threads.

This differed from the Google-internal implementation (which waits for all work to complete in ~UnboundedWorkQueue()) and caused a lock-order inversion deadlock when an in-flight worker thread called UnboundedWorkQueue::Schedule() concurrently with ~UnboundedWorkQueue() (e.g., in //xla/pjrt/c:pjrt_c_api_gpu_test_nvgpu_any during PjrtCApiTest.BufferTransferImmutableUntilTransferCompletes, where CopyRawHostToDeviceAndReturnEvent calls ExecuteWhenReady -> Schedule() on the H2D Dispatch worker thread after on_done_with_host_buffer has already unblocked client destruction):

• Main thread in ~UnboundedWorkQueue() holds thread_pool_mu_ and blocks in pthread_join waiting for the worker thread.
• Worker thread in Schedule() holds work_queue_mu_ and blocks trying to acquire thread_pool_mu_.

Update default/unbounded_work_queue.{h,cc} to track num_running_functions_ (guarded by work_queue_mu_) and wait in ~UnboundedWorkQueue() until work_queue_.empty() && num_running_functions_ == 0 before setting cancelled_ = true and joining thread_pool_. Also update unbounded_work_queue_test.cc to expect all closures to finish in RacyDestructor and add NestedClosureDuringDestructor to cover scheduling nested work during destruction.