Adding ability to record intermediate graph outputs

[agnikumar]

Calibration can be used to improve quantization techniques, adding reduced-precision computation for neural networks while retaining high accuracy without retraining. calibrate.py adds ReduceMin and ReduceMax nodes to all Conv and MatMul nodes in a loaded ONNX model and ensures their outputs are stored as part of the graph output, extracts intermediate output values after inference, and returns a dictionary mapping added node names to average (ReduceMin, ReduceMax) values as input to quantize.py. Example usage:

python calibrate.py --model_path=<'path/to/model.onnx'> --data_set_path=<'path/to/data/folder'> --calib_mode='naive'

[msftclas]

All CLA requirements met.

[jywu-msft]

add some tests?

[askhade]

In our last discussion we decided not to replace the existing code. Looks like your PR is replacing the whole existing script. Can you please make that change? Everything else looks good but I cant merge this till you make this change.... Thanks!