Many thanks to Venemo on #dri-devel for helping out with intrusction level parallelism.

The tiles can now divided in both x and y direction, which allows for:
- Less divergence on typical inputs
- Faster texture stores through coalescing

Performing the normalization in path_coarse.comp reduces redundant
computation as each thread in SIMD computes for a different path.
Gives a slight overall performance boost.

The normal vector direction was flipped to simplify logic.

Inspired by the original kernel4 code, this change avoids the expensive lookup whenever it's determined that the lookup result would be discarded (bit AND against a zero).

It's unfortunate that we end up doing a lot of redundant work in kernel4, but at least we gain up to 10% of improvement for now.