Running block stacking network models on the real robot

[ahundt]

We need to be able to load models from both enas and hyperopt and execute them on the real robot.

1. At multiple time steps for each action step we will run inference to get a regressed pose out
2. Run a classifier to see if the inferred pose is good enough (success/failure classification)
3. if it is predicted successful, move towards that pose
4. repeat until actual robot pose matches regressed pose (< 0.25 cm distance error)
5. execute action step (close/open gripper)

Stacking code is here:
https://github.com/cpaxton/costar_plan/blob/master/ctp_integration/scripts/run.py

[ahundt]

Here is the pull request with this work in progress:
#520