This will hopefully get rid of the flaky tests we've seen in CI. It fixes two issues:
- We weren't guaranteed to get enough cache flushing around copies to the stencil shadow surface
- I botched my code in anv_image.c and the shadow ended up W-tiled which kind-of defeats the point
With these patches, we're down to 16 reliable failures in the group dEQP-VK.renderpass.*.formats.s8_uint and the failures are all self-dep tests where the only real solution would be to emit swizzling code in the shader. We can leave that for a rainy day.