[BUG] OpenImageIO's default thread pool initialization has a data race.

**Describe the bug**

A clear and concise description of what the bug is. What happened, and
what did you expect to happen instead.

OpenImageIO's default thread pool initialization seems to have a data race.
In OpenImageIO 2.4.8 (and it seems to still be the same in upstream main)

```
thread_pool::thread_pool
-> thread_pool::Impl::Impl
-> thread_pool::Impl::resize
-> thread_pool::Impl::set_thread
-> starts a new thread using the lambda f
-> that lambda references thread_pool::Impl::q before thread_pool::Impl's constructor has finished on the main thread
```

Full TSAN error with proprietary parts of the stack removed beloe

**OpenImageIO version and dependencies**

2.4.8

**Evidence**

- Error messages (paste them here exactly)
- Screenshots (if helpful)
- Example input: If the problem only happens with certain image files, please
attach the smallest image you can make that reproduces the problem.

**IF YOU ALREADY HAVE A CODE FIX:** There is no need to file a separate issue,
please just go straight to making a [pull request](https://github.com/AcademySoftwareFoundation/OpenImageIO/pulls).

Start with the thread_pool::thread_pool → thread_pool::Impl::Impl → resize → set_thread call chain described in the issue. Reproduce the initialization under ThreadSanitizer and inspect when the worker lambda accesses Impl::q relative to construction. Done means the reported data race is eliminated and the ThreadSanitizer failure no longer occurs.

Written by the indexing model from the issue text.