Extend the py-torch package to explicitly specify the location of FP16

[koparasy]

Add a simple patch to the cmake file of torch.

Torch when using USE_SYSTEM_FP16 variable assumes the FP16 headers to exist under a hard code path (/usr/include/fp16.h). We patch the cmake to check for an environment variable that sets the path of fp16.h and we set that variable in the torch package.py.

[koparasy]

No, I can do so.

[adamjstewart]

Ping me on the upstream discussion too.

[alalazo]

Closing, since we moved the package repository. Feel free to reopen at / migrate to https://github.com/spack/spack-packages if you want to continue work on this.