Tests fail when no GPU is available

[dvicini]

Hi,

I've never run into this before, but it turns out that when running the tests on a Linux machine without a GPU with the CUDA variant compiled (e.g., using the pip-installed module), we get a bunch of more or less confusing error messages.

Would it make sense to skip CUDA tests if no GPU is available? We have always done this according to the available variants, but maybe an extra check would be reasonable here? This isn't urgent at all, as one can easily just disable the CUDA variant when compiling. I was just surprised to get unclear test errors like Critical Dr.Jit compiler failure: jit_var(r891839089): unknown variable! when running the tests in src/core using the pip-installed Mitsuba

Best, Delio