Proper support for Mochi and temporal vae

[AustinMroz]

Mochi is now available in core, but it's temporal vae introduces some technical difficulties that are not easily solved as the number of frames must be congruent to 1 mod 6.

• Batching vae opperations is difficult since frames in latent space interpolate between one another.
  • More testing is needed, but I suspect that batching will have sideeffects on output even with forced overlap.
• vae encode will truncate frames if not passed a number congruent to 1 mod 6. I don't foresee a sane way to restrict selections in the UI, but am considering adding a hint that a selected number is valid for mochi instead of preventing selection in the widget.
  • Under basic workflows, an invalid frame count is a minor annoyance.
  • Under Meta Batch workflows, this results in skips in the output and audio desync.

Ideally, any solutions would not be tied to mochi in case other temporally aware vae are added in the future.