Add adversarial inference.

[amuamushu]

One thing we want to do is train on clean data (already done for run_pyramid_simple) and run inference on adversarial data (need to do for this issue).

TODOS:

• [x] Add proto field in model or dataset protobuf (this will be used to indicate we want to create adversarial examples out of the image rois)
• [x] Add code for perturbing the image rois and recalling fc_output_layer.build() to get the updated output.
• [x] create config file for adversarial training that includes proto field
• [x] copy over training checkpoints from pyramid_run_simple
• [x] create bash script for running adversarial inference

[amuamushu]

Currently testing out clean model + adversarial inference in https://github.com/amuamushu/adv_avod_ssn/tree/messy-run

[amuamushu]

I realize we don't have to copy over the training checkpoints because the model protobuf has a field for pretrained_ckpt. It was only being used for fine tuning with sin input so I had to modify the code to use it outside that case too.