Support for dynamic int8 quantization

Hi there!

Could we add the functionality to add int8 quantization as a `--precision` option?

I believe the easiest way would probably be using onnxruntime `quantize_dynamic` function after converting to fp32. Not really sure what the difference would be for using MatMulNBitsQuantizer with `bits=8` option.

The reason is that some architecutres (specifically ppc64le) use matrix math optimization for int8 but do not have them for int4.

Would a PR in that direction be accepted?

Start by locating the existing --precision handling and the path that converts models to fp32. Compare the proposed onnxruntime quantize_dynamic approach with MatMulNBitsQuantizer using bits=8. Done means int8 dynamic quantization is available through --precision and supports the ppc64le use case; no file or test is named in the issue.

Written by the indexing model from the issue text.