.cu files should not include torch/extension.h

[lostmsu]

See https://github.com/facebookresearch/pytorch3d/issues/1127#issuecomment-1071206320

This causes nvcc trying to deal with pybind, which fails with

error: too few arguments for template template parameter "Tuple"

[lostmsu]

I tried adding libraries=['cublas'] argument, but it did not seem to affect linker flags at all.

[crcrpar]

Could you confirm that with this change we can build --cuda_ext extension on Linux?

[ksmaze]

Could you confirm if upgrading to use cuda 11.7 resolves the issue? It should be resolved in cuda based on comments: https://github.com/pytorch/pytorch/issues/69460#issuecomment-1278305417

[NiallHornFX]

Likewise, does anyone know if this is resolved with CUDA 11.7 ?