enable building of deepspeed for target GPU's on virtual linux environment without access to real GPUs and /dev/kfd driver.
earlier deepspeed used only the runtime method torch.cuda.is_available() method to check whether to build for rocm/cuda or nvidia cuda, now on rocm case the gpu list can be set as a build parameter. (another patch applied earlier)
fixes: https://github.com/lamikr/rocm_sdk_builder/issues/75