WIP: CI

This is what I'm using atm to check that my patches don't introduce regressions.

Right now it needs a special runner to be associated to the project, as it uses the GPU. We can look at falling back to LLVMPipe if it's not too slow. Currently it takes about 14 min to run deqp gles2 and gles3 on both GL and GLES hosts, plus piglit/gpu on GL hosts.

To be practical, it uses vtest and needs two patches from another MR. This branch also has the HW atomic patches because otherwise some tests fail.

We'll see once it finishes that this MR doesn't pass CI. That's because in master a bunch of dEQP-GLES3.functional.fbo.msaa.8_samples tests have switched to NotSupported. Gert thinks that this is fixed by one of his in-flight series so I haven't debugged.