[MSAA] Blocky artifacts with the even-odd fill rule

[armansito]

I added a new test scene for fill rules in #389. When I tried it with both 16xMSAA and 8xMSAA modes I observed some visual artifacts.

This is how it looks with Analytic AA:

This is with MSAA:

[raphlinus]

This is consistent with the even-odd rule just not being handled correctly in the msaa fine rasterizer, especially as it looks like the logic in coarse is doing the right thing for solid blocks. It wouldn't be hard to add a case, but the best solution is to write a specialized version that uses 1 bit per sample to accumulate winding numbers, using xor. The performance of that should be excellent, with no extra cost to using 16 bits over 8.