Add bf16 Precision Support for TensorRT

[zelenooki87]

I would like to inquire about the possibility of adding bf16 precision support for TensorRT.
As you may know, some ONNX models do not work correctly in fp16, often resulting in black or artifacted images. In contrast, many of the newer super-resolution ONNX models are designed to run in bf16, which is significantly faster than fp32.
Could you please consider adding support for bf16? I believe this would be a valuable feature and hopefully straightforward to implement.
Thank you very much in advance.

[zelenooki87]

@joeyballentine Is there any chance you could implement this? It would be very helpful. For some models that don't currently support fp16, this would provide double the inference speed compared to fp32. I would appreciate a response. Thank you.

[joeyballentine]

I can look into it

[zelenooki87]

@joeyballentine Regarding the other issue, is it possible to get something from it as well? Thank you very much.
#3158

[joeyballentine]

not sure about that one but i can look into it