[Feature Request] Let quantize_dynamic work with float16

Nobody has claimed this yet.

### Describe the feature request

Currently quantize_dynamic() takes an onnx in float32 format an outputs an onnx in int8 format.

It would be nice so that this also worked on an onnx in float16 format.

Also, the new int8 onnx requires inputs of float32. While I would like to put intputs of float16 and it to work only with float16 internally.

### Describe scenario use case

Workflow speed up and possible inference speed up.
Currently my GPU can run float16 faster than float32. So if there is any float32 lurking inside the quantized model this is bad news.
This seems like a pretty easy fix to make. Just to add an options for float16.

Start by locating quantize_dynamic and tracing how float32 inputs become int8; no file or test is named in the issue. Done means float16 ONNX models can be processed and the resulting quantized model accepts float16 inputs without unwanted float32 computation, with coverage for both requested paths.

Written by the indexing model from the issue text.