feat: Add Variable-Q Transform

[gudgud96]

Add VQT feature, similar to librosa's implementation. TLDR:

• Add a gamma to dampen the Q factor at lower frequencies, while maintaining roughly a constant Q at higher frequencies.
• Main difference is that VQT create kernels for each octave based on the new Q, but CQT can pre-compute the kernels

Test plan:

1. plt.imshow VQT magnitude spectrograms between librosa and nnAudio, overall does not have much difference.
   There is still some minor value diff between both, mainly due to lengths - librosa uses float, but nnAudio uses int (I keep this implementation so that it is consistent with nnAudio CQT implementation)

2. Added two tests for VQT: (i) at gamma=0, VQT and CQT should be the same; (ii) keep the value diff between librosa and nnAudio within a baseline range (see tests/test_vqt.py)

3. Ran both tests/test_vqt.py and tests/test_cqt.py, both passes
   

[gudgud96]

@KinWaiCheuk inviting you as reviewer, much thanks!

[KinWaiCheuk]

Isn't this same as line 46? Both self.bins_per_octave and self.filter_scale seem to be unchanged inside the for loop. So maybe there is no need to recompute the Q here? Or am I misunderstanding something?

[gudgud96]

Sorry for replying late. You can understand this Q as an "initial value of Q".
Following librosa's implementation, the "variable" part actually lies in lengths of create_cqt_kernels-

Ours: https://github.com/KinWaiCheuk/nnAudio/pull/113/files#diff-8fea6e5f3e058527d6bbfe52b2dd2d9e756425b37d39bac07541f7f5cb1ccce5R424

Librosa's: https://github.com/librosa/librosa/blob/381efbd684c01ae372220526352d34fd732d3b1d/librosa/filters.py#L802

The gamma will change the filter lengths as compared to CQT. From what I understand, since filter length and Q are interrelated, by changing the lengths we can also view it as "variable-Q".