Optimize atomic floats on nvptx

[taiki-e]

The CPU targets do not have instructions for atomic float, but the GPU targets have these instructions.

• load/store https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-ld
• RMW https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#parallel-synchronization-and-communication-instructions-atom