feat: add spectrum caching method

[rmatif]

Yet another training-free acceleration method. This PR implements Adaptive Spectral Feature Forecasting for Diffusion Sampling Acceleration, currently for UNet only. For DiT models, we already have enough options in my view

This could replace and deprecate ucache, which was only ever an experimental method

Example usage:

/build/bin/sd-cli -m models/model.safetensors -p "a cute cat" --steps 20 -H 1024 -W 1024 --fa -s 42 --cache-mode spectrum --scheduler simple --sampling-method euler

Steps	Baseline	Spectrum
20
	×1.0	×1.82
30
	×1.0	×2.14
40
	×1.0	×2.50

[Green-Sky]

For DiT models, we already have enough options in my view

Personally more interested in the DiT variant. They have some nice speed up and consistency claims.

Anyway, great work. :)

[GreenShadows]

Based on what is claimed in these papers, this method is likely to be more efficient in terms of speed and quality than current DITs.

[JohnLoveJoy]

Impressive how the quality is preserved in the 20-step example!

[wbruna]

currently for UNet only. For DiT models, we already have enough options in my view

So... If I do request --cache-mode spectrum for Z-Image Turbo, and the log says it managed to skip 2 of the 9 steps, is it actually using another method?

[INFO ] stable-diffusion.cpp:3488 - generating image: 1/1 - seed 42
[INFO ] stable-diffusion.cpp:1812 - Spectrum enabled - w: 0.40, m: 3, lam: 1.00, window: 2, flex: 0.50, warmup: 4, stop: 90%
[DEBUG] ggml_extend.hpp:1772 - z_image compute buffer size: 548.22 MB(VRAM)
  |==================================================| 9/9 - 4.67s/it
[INFO ] stable-diffusion.cpp:2333 - Spectrum skipped 2/9 steps (1.29x estimated speedup)
[INFO ] stable-diffusion.cpp:3530 - sampling completed, taking 42.07s