Cutlass support compatibility with nv_bfloat

[YoniSp]

Hi I am trying to create matrix and multiply by using nv_bfloat without success.
There is an option to use nv_bfloat or only using cutlass::bfloat16_t is valid?

[hwu36]

cutlass::bfloat16_t can be used in both host and device. It is also compatible with nv_bfloat. So, use cutlass::bfloat16_t whenever possible. Use reinterprect_cast<> whenever you need to switch.