feat: parallel scan extension for CPU

[yoyolicoris]

Accelerate first-order filter on CPU for small batch sizes.
This PR serves as a first step to gradually remove numba code and use the native cpp bindings by PyTorch. #10 could be merged after this PR.
The empirical benchmark shows the extension is faster when batch size < 2.

I plan to keep the main for development purposes, so the version will always be *.dev. I will make a separate branch for the stable version.

[yoyolicoris]

A separate function like first_order_recurrence(x, a, zi) might be clearer to users about this update.

[christhetree]

I left some comments, otherwise LGTM. Also I don't think I understand what Recurrence is doing but TBF I didn't dig deeper into that part of the code.