feat: ancestral sampler implementations for flow models

[wbruna]

Euler A, 8 steps	DPM++2S A, 5 steps

On top of #1364 because it'll conflict with it anyway, and the eta parameter is useful for testing.

[leejet]

Could you provide some theoretical explanation? That would help me review the related changes more effectively.

[wbruna]

Could you provide some theoretical explanation? That would help me review the related changes more effectively.

This actually came from an observation that "Euler A" worked fine in ComfyUI, but not on sd.cpp. Looking for a reference implementation, I found https://github.com/capitan01R/ComfyUI-ZImageTurbo-FlowSampler , and reimplemented it with adjusts for better numerical stability; but it produced blurry / oversaturated results.

Looking at the ComfyUI calculation, it includes an additional ratio that seems to work much better: Comfy-Org/ComfyUI@d605677 . The problem is that it doesn't offer much explanation about where it came from.

I've rebased the Euler A implementation for now, with the derivation of the expressions I used, and isolating the additional ratio (it's enough to replace it with 1.0 to change it into the algorithm from the first reference).