Fix version comparison for defining CUDA11OrLater
Before this PR `CUDA11OrLater` was incorrectly set to `False` when `torch.version.cuda == "11.0"`.
`torch.version.cuda` returns major and minor CUDA versions, it doesn't return patch info.
LooseVersion comparison was calling `[11, 0] >= [11, 0, 0]` which evaluates to `False`.
