Proof of concept fix for Flax LCM PR

[pcuenca]

Splitting the key makes the denoising loop work without the artifacts previously observed.

This is not meant to be the final solution, but a proof-of-concept demonstration. I'm not keen on introducing RNG Key management at the pipeline level; instead, I wonder if we could set the key as part of the scheduler state and have the scheduler manage it. In any case, I suggest to merge this in the in-progress PR and move the discussion there.

Note also that I believe the solution in FlaxDDPMScheduler is wrong, as it will always generate the same random variance every time the function is called. We do need to carry the RNG state.

I have tested with DEBUG True and False, but without jitting the pipeline.

(Commit 7e0a0cd is unrelated, sorry; it's already in main huggingface#6272)