ROCm 6.3.2 rebuild

Patches must be applied before `make generate`!

Fix typo in AMD gpu targets. Now, the library includes support for all listed targets, not just gfx906.

Fix GPU target handling. Fixes python-pytorch#10

ROCm 6.0.2 rebuild

Switch to git sources as ROCm 6 requires upstream patches that are not yet part of an official release. This requires us
to generate the GPU code manually in prepare(). Furthermore, CXXFLAGS has to be modified to include the current ROCm
version.

Change install dirs to CUDA and ROCm paths to be able to install them side by side.

I had to disable magma-sparse for the time being but that's ok as pytorch doesn't need it.