more fixes

Support totally random voices (and make fixes to previous changes)

Add support for extracting and feeding conditioning latents directly into the model

Move everything into the tortoise/ subdirectory