Downcasting out of range should clamp for float => 8b/16b

[p0nce]

Else the wraparound makes spurious bugs to fix, it also break expectations. Remove < 0 and > 1 Basically all scanline function with float -> int should be modified.