We now use a fixed-size kernel (right now 17), for each mip-level.
And we workout the mapping from roughness to such mip.
This improves performance by reducing significantly the number of taps
per mip, however, we don't control how much blur is actually applied
at each level. This gives a mip-chain that is proportional to
'roughness'.
It's now easier to tune quality vs. speed by tweaking the kernel's width
as well as how we map 'sigma' to a the kernel width.
60b664d3b9