Fix ThreadPool shutdown deadlock on Windows with CUDA

[Deverydoo]

Summary

Fixes application hang when destroying CTranslate2 models on Windows with CUDA GPU acceleration. The ThreadPool destructor deadlocks because worker threads get stuck in blocking CUDA synchronization calls during shutdown.

Problem

When a CTranslate2 model is destroyed (e.g., after Whisper transcription or NLLB translation completes), the ThreadPool destructor calls queue.close() then worker->join(). A race condition causes deadlock:

1. Worker thread enters idle() → synchronize_stream() (blocking CUDA call)
2. Main thread calls queue.close() → sets _request_end, notifies condition variable
3. Worker is blocked inside synchronize_stream(), not waiting on the condition variable — notification is lost
4. Worker::join() blocks indefinitely waiting for the worker to exit

This affects any application using CTranslate2 with CUDA on Windows. Confirmed on:

• NVIDIA RTX 4090 (Ada Lovelace, sm_89) — Windows 10, CUDA 12.8 and 13.2
• NVIDIA RTX 5070 Laptop (Blackwell, sm_120) — Windows 11, CUDA 13.2

Related issues: #1782, SYSTRAN/faster-whisper#71

Changes

include/ctranslate2/thread_pool.h:

• Worker::join() now accepts timeout_ms parameter (default 5000ms). If the worker doesn't finish in time, both threads are detached to prevent indefinite blocking.
• Added Worker::prepare_shutdown() virtual method, called before queue.close().

include/ctranslate2/replica_pool.h:

• ReplicaWorker overrides prepare_shutdown() to set _shutting_down atomic flag with release semantics.
• idle() checks the flag with acquire semantics before calling synchronize_stream(), preventing the race.
• finalize() also sets the flag before resetting the replica.
• Added std::atomic<bool> _shutting_down member.

src/thread_pool.cc:

• ThreadPool::~ThreadPool() calls prepare_shutdown() on all workers BEFORE queue.close().
• JobQueue::get() releases the mutex before calling before_wait() callback and checks _request_end to avoid calling synchronize_stream() during shutdown.
• Worker::join() implements timed join using std::promise/std::future with detach fallback.

Test plan

• Whisper large-v3-turbo transcription on RTX 4090 (CUDA 13.2, Windows 10)
• NLLB-200 translation after transcription on RTX 4090
• Whisper transcription on RTX 5070 Blackwell (CUDA 13.2, Windows 11)
• Application exits cleanly after transcription + translation
• Multiple sequential transcription runs without hang
• No regression in inference accuracy or performance

[Purfview]

There is another PR for the same issue: #1912

[3manifold]

I think it's time to replace third_party/BS_thread_pool_light.hpp, include/ctranslate2/thread_pool.h etc. with the latest version of https://github.com/bshoshany/thread-pool (where the (deprecated) BS_thread_pool_light.hpp idea started in the first place).