CUDA: Lower minimum device architecture check to sm_60.

Native FP16 packet math intrinsics (e.g. __hfma2, h2exp, h2log) only
require `__CUDA_ARCH__ >= 530`, and Pascal (sm_60) has full native FP16
hardware support and declares all needed intrinsics.

Relaxing the architecture guard from sm_70 to sm_60 allows targeting
Pascal GPUs (e.g. Tesla P100) while continuing to guard against
pre-Pascal architectures that lack native half2 intrinsics.

This is Google's default GPU test environment.