shaders/sampling: enable antiringing for EWA scalers (V2)

Based on two-pass design where the main interpolation kernel is combined with a smaller, second gaussian (non-ringing) filter, with high exponent (hard-coded as 8 for performance/quality reasons) to bias it towards low/high values respectively.

This is unfortunately rather slow (about 50% slowdown), but much, much higher quality than all other approaches I came across. It may be possible to optimize it slightly more, (or just drop the numbers to make it faster at the cost of quality), but for now, this represents a vast quality improvement over the status quo of no antiringing.

Supersedes !504 (closed)