Reconsider lavapipe

[jimblandy]

The wgpu sources have various comments here and there remarking on how lavapipe doesn't implement Vulkan well enough to fully support WebGPU, but I wonder if that's true any more. Apparently Linux Mesa lavapipe 22.1.2 passes the Vulkan 1.3 CTS.

We should go through wgpu and see what works now and which caveats still apply. It would sure be nice to have lavapipe for software-based testing.

[teoxoy]

The zero_init_workgroup_mem test failing in CI (running via swiftshader currently) is becoming more and more prevalent. I wonder how lavapipe compares.

[cwfitzgerald]

The zero_init_workgroup_mem test failing in CI is our fault not swiftshader's, it's the same thing affecting WARP: #3193

This wasn't obvious before as the vulkan validation layers were off. Afaict, there have been no spurious failures from swiftshader (that weren't our fault), but there have been quite a few from lavapipe. We should definitely try it again, but I have a lot more confidence in swiftshader than lavapipe.