tests: Handle a few test failures on native NVIDIA.
Specifically, on my particular version of Windows 11 and the NVIDIA driver, and on the NVIDIA GeForce 4060 Mobile that I happen to own. Unfortunately I don't have many other NVIDIA GPUs to compare with.
- Patches 1/6 and 3/6 just look like NVIDIA driver bugs, I can't of any way to explain the results I'm seeing. So I think the most sensible thing to do is to just ignore what happens there. Do we want to introduce some equivalent of
bug()for native drivers, so that they at least leave a trace in the logs? - Patch 2/6 is a bit puzzling, a couple of probes expect to find a non-zero value in a component that is statically written zero to. I suppose it wasn't appropriately tested when it was introduced, with my changes it passes on NVIDIA, AMD and WARP.
- Patch 4/6 implicitly relies on unspecified behavior, now is fully specified.
- Patch 5/6 might be in the same league as 1/6 and 3/6, but the correct behavior is less clear here. We could just decide that no consistent behavior is observed for SM<4 and remove the two tests altogether.
- Patch 6/6 is the usual NVIDIA being super lazy on trigonometric functions.