Refactor F.lfilter and F.*_biquad autograd tests

[yoyolicoris]

Adopt the autograd test model from transforms/autograd_test_impl.py.
Also add more combinations of testing parameters.

[mthrok]

Thanks for the PR.

Although I like the thorough testing, it's hitting the CI job time out limit.
Can we reduce the time it requires for the test somehow?
Autograd tests are by nature expensive itself.

I think it's good enough to use only one value for central_freq and Q.
I could be wrong but it feels that 2, 100, 800, 4000, 8000 are in a safe range for gradient computation.
If it is around 0, I would be more cautious. (but again, this is my intuition and not backed by any hard fact. So if you disagree, please let me know)

Things like const_skirt_gain should be parameterized because that change the code execution path, and we want to cover all the path.

[yoyolicoris]

I think it's good enough to use only one value for central_freq and Q.
I could be wrong but it feels that 2, 100, 800, 4000, 8000 are in a safe range for gradient computation.
If it is around 0, I would be more cautious. (but again, this is my intuition and not backed by any hard fact. So if you disagree, please let me know)

Sounds reasonable, I agree.

[mthrok]

Thanks!