AltiVec gfx/ycbcr has some inefficiencies

[classilla]

• Loading from the YUV table is not 16-byte aligned, and we load a LOT of data.
• The unaligned stores are probably not as efficient as they could be.

[classilla]

Actually, the unaligned stores are probably optimal since we have no alignment guarantees in the original x86. Even switching to double could hurt, since the doubles may not be 8-byte aligned, and would cause a hardware fault. But there are still a couple other things to do.

[classilla]

This is notably better and I'm not sure we can improve further, so let's call it good.