docs: add MolmoAct2, FastWAM and VLA-JEPA

[da-dawit]

3 New Models Added:

• List the three policies in Supported Models and add a training command for each, taken from the LeRobot docs.
• MolmoAct2 is the one policy with a separate learning rate per component.
• FastWAM concatenates the cameras into one image, so mixed camera resolutions need no extra flag.
• VLA-JEPA: freeze_qwen trains only the action head, and resize_images_to puts every camera at one resolution.
• Extend the model picker with a branch for the three.

Also correct two points on the existing policies:

• Fine-tuning SmolVLA from its base checkpoint needs --rename_map as well, since the base declares camera1/2/3 while XVLA declares image/image2/image3.
• Diffusion requires every camera to share one resolution, so a robot with mixed camera sizes needs a single camera selected.