Add more dpmsolver++2m samplers to solve SDXL sampling artifacts

[rockerBOO]

Visual artifacting for sampling using dpmsolver++. This adds more samplers but dpmsolver++_2m_lu should produce better results on SDXL.

huggingface/diffusers#5433
huggingface/diffusers#5541

[rockerBOO]

I thought I did this correctly as referred to in the linked issues, but there are a few things I need to figure out. I can't easily test this outside of training so will take some time to fiddle with it. The schedulers are having a set value being passed but not sure why. I looked back in the history but did not find any indication of why, except it is using the defaults from one of the schedulers.

    scheduler = scheduler_cls(
        num_train_timesteps=SCHEDULER_TIMESTEPS,
        beta_start=SCHEDULER_LINEAR_START,
        beta_end=SCHEDULER_LINEAR_END,
        beta_schedule=SCHEDLER_SCHEDULE,
       ...

So I am trying some experiments with it removed but it will cause samples to be different than before with the same settings, which can make it difficult to compare. At least find something that works better for SDXL sampling and then we can work back from there.