feat: add LCM scheduler and use it by default for LCM sampling

[wbruna]

Implement the linear LCM scheduler, as defined e.g. in Diffusers' LCMScheduler, and select it by default with the LCM sampler.

The main change is ensuring the timesteps match the LCM training schedule. The default discrete scheduler can choose unmatched timesteps, which tend to require higher step counts to avoid artifacts and blurring. It is more noticeable with detailed images, or when the model isn't pushed too hard (e.g. with minimal prompts). simple and sgm_uniform tend to generate much better results, but I’d guess that’s mostly by chance.

Not every generation is better, and when compared with simple, it isn’t clear there’s even a benefit; but it’s easier to notice the difference when comparing with discrete. A few examples with 4 steps, somewhat cherry‑picked to avoid the worst abominations:

discrete	lcm

I can’t properly test video generation, so I didn’t change its default; just let me know if I should.

Another consideration: since this change mostly affects how the timesteps are rounded, a better alternative might be adding an extra parameter to all schedulers to request this behavior. This could make LCM models (and perhaps other distillations) behave better across all schedulers at once.