Numerous enhancements and fixes (Multichannel audio support, cleaner predictions, scalable losses)

[pokepress]

Admittedly, this is a pretty big update. I'll break down the changes:

• Adds support for multichannel audio
• Makes most losses scalable
• Adds DC offset loss (DC offset measures how evenly a sound wave is centered around the 0 line-I sometimes had cases with seanet where the waves would become un-centered, and having this loss helps track it)
• When using predict on an audio file, borders between segments are crossfaded to make transitions less abrupt
• Enumerates some lists before looping (seems to improve performance)
• Adds support for bfloat16 (on modern GPUs, tends to be faster and use less memory)
• Adds configurations for 44.1->44.1 audio cleanup
• Updates requirements file to newer versions of components
• Allows saving best models without using cross validation
• Fixes some subscript names
• Fixes some typos
• Removes unused parameter from predict logic