add multi-pass inference code

[saum7800]

The code changes facilitate multi-pass inference: passing an image with the one hot labels encoded into the image in a different pixel each time and choosing the predicted label as the image that yielded maximum sum of squared activations.

• changed data loading to also load samples of images with each image having the label encoded in a different pixel (from first 10 pixels)
• added function to compute the multi-pass inference label
• added code in main.py to run the new inference during validate_or_test
• added a config variable to run the code without needing to do this inference