Consistent and flexible configuration for GPU acceleration

[MedericFourmy]

We need to offer a more consistent way to define which type of hardware acceleration devices happypose runs on. Here are the current and desired behaviors I have in mind so far:

• Choose whether torch runs on cpu on gpu
  • current: use cuda if available, use cpu otherwise
  • desired: user should be able define the preferred acceleration device (e.g. through environment variable)
• Choose whether to run the renderers on gpu or cpu.
  • current: for pybullet, the choice is possible. For panda3d, use GPU if cuda is available.
  • desired: configurable for both
• AMD ROCm support: Some computing cluster (e.g. LUMI) require the use of AMD GPUs. ardware acceleration through AMD ROCm (~ equivalent of Nvidia Cuda) should be supported by latest torch versions. The refactoring should take this into account.

EDIT: ROCm support seems to "masquarade" itself as cuda so there should not be any adaptation in this regard.
See this gist