Disable gradients for null conditioning when CFG is enabled

[lucasnewman]

I had a training run loss blow up to NaN after a while with conditional drop enabled, and it looks like the null conditioning had gradients enabled and would eventually overflow from debugging the model weights. This just disables gradients for the null conditioning parameter. I verified the loss still converges as expected.

I also included a couple of drive-by fixes (let me know if you want them in another PR):

• Fixed a couple of undefined variable references
• Pinned the fractional mask tensor to the accelerator device
• Added an option to decode to EnCodec quantized codes instead of audio / latents, which allows using e.g. the multi-band diffusion decoder from Meta.

[lucidrains]

@lucasnewman that's really interesting learned null conditioning led to instability! i'll have to think about that one

rest lgtm!

[lucidrains]

may be seeing synergy between gateloop and attention layers (combined green run actually has less parameters than either the gateloop or attention run alone)

recommend giving that a try!