Could we increase CPU-bound performance by manually managing the GIL?

I’ve done some research on this for Cubed: https://github.com/cubed-dev/cubed/issues/497

I’m curious that if it would be worth it to add a Cython build step to Xee. From SO and docs that I’ve read, there may be more CPU performance we can squeeze out if we manually manage the GIL when interacting with concurrent ThreadPools.

I suspect @shoyer may know if this is an issue, but measuring is always better than guessing.