Discard penultimate sigma for DPM-Solver++(2M) SDE

[catboxanon]

Description

Based on discussion linked below, it is desirable to have the new DPM-Solver++(2M) SDE sampler discard the penultimate sigma.

https://gist.github.com/crowsonkb/3ed16fba35c73ece7cf4b9a2095f2b78?permalink_comment_id=4568060#gistcomment-4568060

As an aside, the author seems to recommend using Karras over the normal scheduler in general. Maybe we could only modify the normal schedule to discard to penultimate sigma by default, but I leave that up to (you).

DPM++ 2M solvers really don't like the normal VP noise schedule, its second derivative (in terms of log sigma) is large at the end. The most friendly schedule to it is exponential, which is evenly spaced in log sigma (second derivative is zero everywhere). Karras is closer to exponential and in my opinion it is close enough to work well.

Screenshots/videos:

(Taken directly from the aforementioned discussion)

Without change

With change

Checklist:

• I have read contributing wiki page
• I have performed a self-review of my own code
• My code follows the style guidelines
• My code passes tests

[Sakura-Luna]

I'm in favor of only modifying the 2M SDE, but adding an option would be nice too.

[Cyberbeing]

I did some quick tests, and I think until a superior solution is found, it would be better not to force discard the penultimate sigma = True for the Karras variant, as only the non-Karras variant seems almost guaranteed to produce completely distorted junk images without it.

My first impression of the 2M SDE Karras variant is that discard the penultimate sigma = True mainly just makes the output different, but on complex scenes the overall image composition can change significantly which can randomly result in discard the penultimate sigma = False ending up with subjectively superior or more prompt accurate output. Hence why I think the Karras variant would best be left with user togglable penultimate sigma option similar to other samplers.

I also think it would be useful if the forced discard the penultimate sigma = True was added to infotext, at least for now, since otherwise it will create a real mess for prompt reproducibility in the future if the change ever gets reverted.