Redundant / multi-phase CQT basis

As noted in #341, the CQT basis can be lossy for high-frequency elements when a large hop-length is used.

Schoerkhuber & Klapuri circumvent this by using a redundant basis where the same kernel is replicated several times at different phase offsets. This gives good preservation of transients, and is generally necessary for high-quality inversion #165 .

Implementing this will require several changes to the code, basically anywhere a CQT is touched. This includes:

• filters.constant_q and friends
• filters.cq_to_chroma
• core.cqt, hybrid, pseudo
• feature.chroma_cqt and friends
• core.cqt_frequencies
• display.specshow
• … and possibly others

[bmcfee]

I’d rather handle VQT as a separate issue.

[bmcfee]

I would not miss anything if LibROSA only supported that.

Good. 😃 Ragged time-sampling is a practical nightmare and I have no interest in implementing it.

Though I would like it to be clearly documented that for lower frequency bands there are essentially duplicated frames (in order to match the higher frequencies’ frame rate). This seems like it could trip up some applications.

The way I see it, it’s better if we just say that all frequency bands are replicated k times, not just the high-F bands where it’s strictly necessary. It’s a bit inefficient, but it simplifies the logic considerably.