Improve DDIM sampling pipeline (support for predict x0 and uneven jumps)

[lukovnikov]

• support for predict_epsilon = False in DDIM sampler
• changed timestep selection such that it is more uniform if number of sampling steps doesn't cleanly divide number of training steps (this leads to not evenly spaced jumps)
• changed implementation of step() to use pairs (x_tm1, x_t) in order to support the previous point and possibly non-linear substep selections (like the quadratic one from the appendix of the DDIM paper) - however, the method still accepts old integers and in that case computes the step size as before.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[lukovnikov]

closed in favor of #1457