Port to HIP / AMDGPU

[tristanitschner]

Hi there,

this is an early RFC and not at all in merge-able state.

I have "ported" your code to HIP and successfully trained your model on a Radeon 9070.

Important points:

• unused (maybe HIP only?) header removal
• adapt chevron syntax, since hipify cannot deal with spaces between the chevrons
• the submodules were adapted based on the most recent master branch, so there are functional changes
• add a sample env.sh file for local python deployment without virtual environment
• adding missing includes / defines in the setup.py files in the submodules
• there are still a lot of warnings when compiling for RDNA, also currently it complies for all architectures, which takes quite some time
• besides training, I didn't test anything else, so there might be issues

If you're interested, please tell me what is the best way forward.