float16 quantization

[JesseFarebro]

Describe the problem or feature request

It would be useful to support quantizing to float16 over the wire and then dequantizing back to float32 on the client.
I'm working with a model where the 2-byte affine quantization in TFJS performs very poorly but using float16 quantization has a negligible impact on performance.

[pyu10055]

Thanks, that is very a good suggestion. @JesseFarebro will you be interested in contributing your code to enable flat16 quantization?

[JesseFarebro]

I'll probably end up putting together a PR for this in the next week or so as I'm going to end up needing this.

[pyu10055]

@JesseFarebro that would be great, thank you!