Crepe

This release adds API (`verbose=`) and command line (`--quiet`) support for silencing CREPE messages (e.g. keras progress bar) during prediction.

This release adds two features where you can specify

- the step size to use (18) for performing pitch estimation: `--step-size`, and
- the model capacity (20) to run pitch estimation with a smaller model `--model-capacity`.

Refer to the command line help `crepe --help` for detailed usage.

These features provides a trade of between the computation time versus the time resolution or slightly lower accuracy, and it can be especially useful when running CREPE on CPU (19).

Here are some ballpark numbers for running CREPE on a 54.31s-long WAV file with various parameters, showing you can get up to 20 times at the cost of slightly lower time resolution and accuracy:

| | full (default) | large | medium | small | tiny |
| -------- | ---- | ------- | ----------- | ------- | ---- |
| **10 ms (default)** | 91 s | 60 s | 35 s | 15 s | 9.1 s |
| **20 ms** | 47 s | 31 s | 19 s | 9 s | 6 s |
| **50 ms** | 20 s | 14 s | 9 s | 6 s | 4 s |

The above are the `real` time reported by `time CUDA_VISIBLE_DEVICES=-1 crepe -s {step} -c {capacity} audio.wav`, on a i7-6600K CPU.

This release changes the behavior of CREPE such that by default frames are **centered** on their timestamps instead of beginning at their timestamps. This way CREPE's default behavior matches that of other audio processing libraries such as Essentia and Librosa, to avoid output misalignment.

This release fixes a bug in that caused crepe to fail (keras model not loaded) when used inside python with via crepe.predcit() or crepe.get_activation()

This release adds support for using crepe as a python module