Add basic mimpmap downscaling support

[jermy]

This is small patch to fix quality issues when scaling images down. It'll create a number of mipmap/half-scaled versions of a given source as required and use the next largest version as the source for bilinear or bicubic scaling.

This uses the same downsampling functions as Skia, but similarly doesn't run SIMD versions of these. As they are integer-based, these should be reasonably quick.

This addresses issue #13, but probably wants some work and tests before merging.

[RazrFalcon]

Thanks a lot 🎉

[waywardmonkeys]

Looking pretty good!

Any thoughts on adding a couple of tests?

[jermy]

@waywardmonkeys any idea of what sort of tests would be useful in this case? I'm imagining just integration tests, but if there are sensible unit tests I could add those? (I think compute_required_levels is the only non-complicated function).

Integration tests would involve some source and expected PNGs so might be a little tricky to keep the size down, but I'm guessing would include:

• That a resized image looks like the same as it currently does if FilterQuality::Nearest is used
• That a "better" image is produced if FilterQuality::Bicubic or FilterQuality::Bilinear are used
• That a 1x1 image scaled down doesn't cause a crash
• That all of the downsample functions are used - I think a source image of 21x19 scaled down to less than 1x1 (ie. 2% size) would be the smallest source image to invoke all of these in one pass

[DJMcNab]

I agree that this needs tests. One note: for mipmaps, you can cleanly pack all the "expected" results into a single file with the same dimensions as the original image, which I think is worth doing.

(That is, I think it's potentially worth having "unit tests" of the actual downscaling)

[DJMcNab]

We should remove this. It's not actually meaningful if it's not in the license, and so can only add extra confusion