Update voxel2clip training for text and resnet

[jimgoo]

• Add argparse instead of the old brittle configurator, all defaults are the same as before
• Support resnet RN50 as the clip variant for testing the MindReader f_mr version
• Flip the forward and backward top 1 retrieval scores so that they match Evaluation.ipynb
• Cast trial to numpy before using it to get the annotations
• Update remote_data=True for average splits so I can run on my own machine
• Make it save the best checkpoint
• Remove unused imports, splitter methods, and other junk things