docs: design SO-101 red-blue joint training

[Liuyuchen0224]

What changed

• Added a design specification for jointly training one OpenPI π0 LoRA model on the local SO-101 red-cube and blue-cube LeRobot datasets.
• Defined multi-dataset loading, per-sample task prompts, SO-101 camera/state/action mappings, and normalization flow.
• Defined delta actions for the first five joints while keeping the gripper action absolute.
• Set the initial low-memory training parameters, including batch size 1 and checkpoint saves every 2,500 steps.

Why

The two datasets need to train one task-conditioned model without physically rewriting either source dataset. The design establishes the smallest compatible OpenPI changes and explicit verification criteria before implementation.

Impact

This PR contains documentation only. It does not change runtime behavior.

Validation

• Confirmed the two datasets are LeRobot v2.0 SO-101 datasets with six-dimensional actions and front/wrist cameras.
• Confirmed the design uses one model, a [true, true, true, true, true, false] delta mask, and a 2,500-step save interval.
• Ran git diff --cached --check before committing.