determine dtype based on gpu compute capability

[zhuangqh]

Is your feature request related to a problem? Please describe.

use different dtype for different generation gpu to achieve better performance.
regarding V100, its compute capability is 7.0, which means it does not support bfloat16 precision.
if we use --torch_dtype=bfloat16, pytorch will support bf16 by software emulation with lower performance.

Describe the solution you'd like

setting dtype value from controller based on gpu compute capability.

Describe alternatives you've considered

Additional context

pytorch/pytorch#124996