Add workaround for nvcc header dependecies bug (#62550)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/62550
I noticed that running the build twice in a row resulted in ~80 CUDA files being
rebuilt. Running `ninja -d explain` shows
```
ninja explain: TH/generic/THStorage.h is dirty
ninja explain: TH/generic/THStorageCopy.h is dirty
ninja explain: THC/generic/THCStorage.h is dirty
ninja explain: THC/generic/THCStorageCopy.h is dirty
ninja explain: TH/generic/THTensor.h is dirty
ninja explain: THC/generic/THCTensor.h is dirty
ninja explain: THC/generic/THCTensorCopy.h is dirty
ninja explain: THC/generic/THCTensorMath.h is dirty
ninja explain: THC/generic/THCTensorMathMagma.h is dirty
ninja explain: THC/generic/THCTensorMathPairwise.h is dirty
ninja explain: THC/generic/THCTensorScatterGather.h is dirty
```
considering `ninja` is working relative to the `build` folder, these files don't
actually exist. I traced this back to the output of `nvcc -MD` containing
paths relative to the include directory, instead of being absolute.
This adds a little script to launch the compiler then resolve any relative paths
in the `.d` file before `ninja` looks at it. To use it, I run the build with
```
export CMAKE_CUDA_COMPILER_LAUNCHER="python;`pwd`/tools/nvcc_fix_deps.py;ccache"
```
There are some possible pit-falls here. The same relative path might work for
two include directories, and the compiler could pick a different one. Or,
the compiler might have additional implicit include directories that are needed
to resolve the path. However, this has worked perfectly in my testing and it's
completely opt-in so should be fine.
Test Plan: Imported from OSS
Reviewed By: ejguan
Differential Revision: D31503351
Pulled By: malfet
fbshipit-source-id: b184c4526679d976b93829b5715cafcb1c7db2ae
