Now when source and destination have non-transparent pixels
we use straightforward OVER algorithm that does expected things:

1) Premultiplies src and dst
2) Blends them using opacity and mask
3) Unmultiplies the value to write into the destination

This algo is even more complex than the one used in COMPOSIRTE_OVER
(yes, we use approximation) and works about 50% slower than OVER.

But since most of the time COMPOSITE_COPY is applied without mask the
slowdown will be hardly seen.

BUG:333227