feat: new IIR backend `ssm` for faster biquads on GPU

[yoyolicoris]

As the title says. Note that the performance gain on GPU needs torchlpc>=0.5 which includes the parallel scan numba kernel. A test function is added to verify lfilter and ssm give the same results.

Notes

• I added two files (not committed), tests/__init__.py and tests/processors/__init__.py, to get pytest to work. I'm unsure how it will be when running the tests on CICD.
• clamp=False is hard-coded in lfilter, as I don't think setting it to True makes sense for audio applications, and it also enables comparing its results with other backends. A historical feature added in early torchaudio to solve the numerical instability of the direct form filter...
• How are the doc pages deployed? Should I commit my locally built doc pages to this PR or do something else?

[sh-lee97]

Thank you for the PR! We merged it.

• We are fixing the import issues, both in the local and the GitHub workflows.
• As we merged your PR, the clamp is now set to False.
• Currently, the documents are run and then committed manually, usually before releasing the commit as a separate version. We are not completely sure, but we might automate this (and maintain the docs per version).