add scaled_linear schedule in PNDM and DDPM

[patil-suraj]

This PR adds scaled_linear schedule in PNDM and DDPM schedulers so those could be used with latent/stable diffusion models.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[anton-l]

LGTM! Also, we should probably take these outside the inits later to avoid copying schedules 😅

[damian0815]

hi, sorry to dig up an old issue. i'm hacking around on Everydream (a dreambooth-like finetuner for Stable Diffusion). i'm noticing substantial differences in outputs depending on whether the beta_scheduler is left at scaled_linear (loaded from the SD2.1 diffusers-format checkpoint) or overridden to linear. in particular, training with scaled_linear makes changes in learning rate have more wildly unpredictable effects on training losses than when it's overridden with linear. i'm not sure how to find out which is more "correct" - would greatly appreciate any thoughts you might have.

[patrickvonplaten]

cc @patil-suraj and @pcuenca herer maybe?

[damian0815]

ahh, nm, it doesn't seem to make that much difference, i was being confused by another hyperparameter i didn't notice was changing. sorry for the noise!