Add convert_float_to_bfloat16 function to avoid overflow

[vraintanakarei]

I used convert_float_to_float16 function to convert fp32 models to fp16, but some models have outputs that have changed significantly due to overflow.
I think bf16 is useful to avoid overflow in such cases.

In the convert_float_to_float16 function, FLOAT16 is hard-coded and could not be changed to BFLOAT16 and so on.
I think that it would be more useful if it could be changed.

Thank you.