Add CLI renderer, optimize onnxruntime

[kcoul]

Submitting this as a draft PR since there are probably a lot of changes within it you don't want to pick up, so best not to merge it directly and instead pick out what you like.

I made a CLI converter target so that I could easily time the conversion speed and see if there were any easy optimizations. There were a few on the onnxruntime side, for RTNeural I'd have to take a closer look.

If you have a "known good" audio file and corresponding MIDI output, along with what parameters you use to convert the file, I could use it as a ground truth while attempting more aggressive optimizations.. maybe Jatin would have some ideas if it were possible to bump to RTNeural v2!