Add ROCm detection to `--torch-backend`

[charliermarsh]

For inspiration: https://github.com/pytorch/pytorch/blob/7cf38d2a0521c5ca292a720bce00617be749a0e7/.github/actions/setup-rocm/action.yml#L24-L49

[charliermarsh]

>>> import torch, pprint
>>> torch.__version__
'2.5.1+rocm6.2'
>>> pprint.pp(torch.cuda.get_arch_list())
['gfx900', 'gfx906', 'gfx908', 'gfx90a', 'gfx1030', 'gfx1100', 'gfx942']

[charliermarsh]

>>> torch.__version__
'2.7.1+rocm6.3'
>>> pprint.pp(torch.cuda.get_arch_list())
['gfx900',
 'gfx906',
 'gfx908',
 'gfx90a',
 'gfx942',
 'gfx1030',
 'gfx1100',
 'gfx1101',
 'gfx1102',
 'gfx1200',
 'gfx1201']

[charliermarsh]

>>> torch.__version__
'2.6.0+rocm6.1'
>>> pprint.pp(torch.cuda.get_arch_list())
['gfx900',
 'gfx906',
 'gfx908',
 'gfx90a',
 'gfx942',
 'gfx1030',
 'gfx1100',
 'gfx1101']

[charliermarsh]

>>> torch.__version__
'2.4.1+rocm6.0'
>>> pprint.pp(torch.cuda.get_arch_list())
['gfx900', 'gfx906', 'gfx908', 'gfx90a', 'gfx1030', 'gfx1100', 'gfx942']

[charliermarsh]

>>> torch.__version__
'2.7.1+rocm6.2.4'
>>> pprint.pp(torch.cuda.get_arch_list())
['gfx900',
 'gfx906',
 'gfx908',
 'gfx90a',
 'gfx942',
 'gfx1030',
 'gfx1100',
 'gfx1101',
 'gfx1102',
 'gfx1200',
 'gfx1201']