Fallback from GPU gen to shared for non-tensor

[jeremylt]

Minor PR to automatically fall back to shared backends when you find an operator with non-tensor bases for GPU

[jeremylt]

When this is approved, I'll make a matching PR to do the same to SYCL

[nbeams]

*-shared doesn't have any non-tensor kernels, only -*ref does (and magma). So if you create a shared fallback, it will actually be calling ref for everything through delegation. I think the fallback (and message) might as well be for ref, to avoid confusion.

[jeremylt]

Good point, I'll update that since its really the same core operator handling between the two