Multithreading: num_threads and no-GIL

[Piezoid]

Adds support for the num_threads parameter and GIL releasing in the Encoder/Decoder.

For batch conversions with ample memory, transcoding multiple files concurrently is often more efficient, but requires reducing or disabling libjxl's parallelism. Releasing the GIL enables concurrent jxl operations in a single python process.

Encoder parallelism can be configured via Image.save(num_threads=<n>). The decoder supports this too, though Pillow lacks a direct way to pass arguments. A settable module constant might be an option—open to feedback on this.

Behavior mirrors cjxl's --num_threads parameter:

-1: machine default, 0: single-threaded.

[Isotr0py]

LGTM! Thanks for your great contribution!