[2Dblock] implement specialized kernels

[januszjah]

In order to make #5338 work even if there are no known values at compile-time we have to make our backend return specialized kernels with optimized version for final case

[januszjah]

I have duct-tape-implementation just for my issue (specific to pitch / width), however this task introduces new feature - specialized kernels, so I'm trying to make it more generalized / reusable.

[januszjah]

I've done version with less changes that can work, but I don't like how assembly looks, because there is a lot of instructions generated that are unused what makes kernel bigger. The problem is that I may rely on SPV DCE too much.

Next steps:
Find a way to put more effort on DCE from SPV side or implement generating of instructions that wrap whole 2d blocks logic in if-blocks that are subject to DCE, however that will make a lot of possibly unnecessary changes.