Add SSAA with SSIM-based perceptual downscaling

[HybridDog]

Super-sampling anti-aliasing with a conventional downscaling algorithm smoothes the image, which is a disadvantage compared to disabled anti-aliasing. The SSIM-based perceptual downscaling algorithm tackles this problem and it parallelises well on the GPU.

The new SSAA method can be enabled with the anti aliasing setting and it is the slowest anti-aliasing technique now.

additional_images.7z.zip

For more pictures and information about the algorithm, see #6976 and the Perceptually Based Downscaling of Images paper.

Closes #6976
Related roadmap goal: 2.1 Rendering/Graphics improvements

How to test

On GNU/Linux with X11, it is possible to render everything shown on the screen with a smaller resolution and upscale it with nearest neighbour scaling by an integer factor: xrandr --output CHANGE_THIS --transform 0.5,0,0,0,0.5,0,0,0,1 --filter nearest (here it's 2x upscaling).
This may help testing the anti-aliasing on a high resolution screen.

In Minetest the new SSAA method can be tested by setting the anti-aliasing setting to antialiasing = ssaa_ssim_based and playing around with different scales (fsaa setting).
By setting gl_FragColor = vec4(ycbcr_to_linear_srgb(col_l), 1.0); in client/shaders/ssaa_ssim_based_3/opengl_fragment.glsl, the algorithm behaves like the current smooth SSAA in Minetest except that it uses gamma-correct scaling.
Note that the new SSAA method also preserves details which may be undesired, for example Moire artifacts if mip mapping is disabled.

Strange behaviour with tonemapping

If tonemapping is enabled, strange artifacts appear with testnodes:generated_png_ck:

With other nodes or without tonemapping, it looks fine.
Perhaps the problem wouldn't appear if the downscaling was applied after tonemapping instead of before.
We can ignore this behaviour because tonemapping in Minetest is a questionable feature anyway as the tonemapping input is already in SDR.

[Zughy]

@HybridDog rebase needed

[HybridDog]

I have rebased it; thanks for the notification.

Recently I found out that there's a patent associated with the paper which belongs to the Percim company I think. I still don't understand what exactly is patented and what this means for this Pull Request. The downscaling method explained in the paper is not a single algorithm but an idea which encompasses many ways to implement a downscaling algorithm.
The patent: https://patentimages.storage.googleapis.com/ad/3d/41/438150ce88bfba/WO2017017584A1.pdf
Percim: https://percim.com/#intro
Comments about the patent: https://gist.github.com/HybridDog/dd95a99d411972f030fb18543280ad60