Make threadpool queue size configurable

[erip]

Analysis here suggests that the queue size is too small to be effective for parallel processing. It's a bit hard to follow the code path, but it seems to be capped at 8*num_threads in the case of LDA. While it isn't obvious what a good default is, this being hard coded seems to be causing issues. It should be a kwarg from infer so users can tune inference.

[erip]

I guess for what it's worth: I wanted to submit a PR for this change, but can't quite get the code in main to build locally on my WSL rig...

[jgb-hda]

For your information: I did manage to compile tomotopy with your proposed changes, however this did not improve the slow inference.

I don't have precise timings or any nice graphs. but it felt as it still took as long as before. I did change 8 to 128. No idea if that is actually a good value, but it didn't seem to change anything anyways, as the CPU was still under high load but not doing much, as it didn't got warm, so I guess it still did spend most of its time waiting …