[Onnxifi] Generic way of passing output shape/type hints (#39229)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/39229
Previously we have a ad-hoc way of passing output shape/type hints which is very limited and doesn't support quantized output. We actually have all the shape_info/qshape_info so we pass them as TensorProto and QTensorProto directly. This will pave the way for us to set output to quantized type in OnnxifiOp.
Test Plan:
```
buck test glow/fb/test:net_runner
```
Reviewed By: hyuen
Differential Revision: D21781515
fbshipit-source-id: dfae3276e8f158eed830f1244bea6420a9135aab