Allow always discarding of penultimate sigma and fix doing 1 less step than specified

[hentailord85ez]

Allows the user to specify if to always discard the penultimate/next-to-last sigma (regardless of the sampler), and makes it so even if you discard the step, you still do the amount of steps specified.

Currently, only the DPM2 (/a) samplers discard the penultimate sigma. However, the quality of other samplers can also be increased by doing the same. This is particularly noticeable if you are using a very low amount of steps (< 10) and are not using a suitably low CFG scale.

With DPM++ 2M Karras (7 steps, 9 CFG scale):

Not-discarding:

Discarding:

Particularly noticeable in the 2nd image across.

Also, currently if discarding the last step, it will sample one less step than the amount of steps specified (which also causes the progress bar to never fill). This is fixed by adding +1 to the steps when getting the sigmas if discarding.

[cluder]

seems to work, its most noticeable as you said with very low steps < 15 or 15, and it seems DPM++ 2M (non-karras) profits a lot from it.
To easier test it, we can add 'always_discard_next_to_last_sigma' to the Quicksettings list

[eaglgenes101]

Are there cases where it's bad to do this discard? (Besides backwards compatibility, maybe)

[hentailord85ez]

Are there cases where it's bad to do this discard? (Besides backwards compatibility, maybe)

If you are already using a reasonable CFG for the number of steps you are using, it can reduce detail in the generation. On larger step sizes, this doesn't matter.

[AUTOMATIC1111]

There is an issue with adding a lot of settings which would make reproduction more difficult since you wouldn't remember if this or that particular setting was on or off, so adding this is not 100% free. Let's see if there is more interest in this before merging in the setting.

[aliencaocao]

Just here to add that I am interested to see this merged