CQT compute downsampling from hop_length

The CQT implementation will currently fail if the hop_length parameter is not an integer multiple of 2**k where k is the number of downsampling operations in the plan derived from fmin and the number of bins.

We should be able to invert that logic so that k is calculated by the number of trailing 0s in hop_length’s binary representation. That way, we can compute cqt at arbitrary hop lengths, and still get the efficiency gain wherever it’s possible, rather than failing.

This is a low-priority issue, but we should get around to it sooner or later.

[bmcfee]

Annoyingly, this will probably trigger another rewrite of icqt, which at present has a hard-wired assumption that the hop length is a multiple of 2^(n_octaves-1). On the plus side, this rewrite would automatically carry over to support vqt #1161, which should be easier now that the vqt implementation can be simplified a bit. I think we can deal with both issues in a single PR.

[cwitkowitz]

Hey @bmcfee. Sorry I’m late to this. It does seem that not scaling gamma would lead to the issues you mentioned. In your last example, what VQT parameters did you use?