fix(qwen-gr00t): stabilize image feature order

[Teeeio]

PR Category

Core

PR Types

Bug Fixes | Improvements | Test Case

PR Description

This PR stabilizes Qwen-GR00T visual input ordering during training.

Previously, the effective image order could implicitly follow dataset-native feature ordering. For LIBERO-style data, this could
produce wrist_image -> image instead of the validated image -> wrist_image order, which changes visual token layout and hurts downstream performance.

Among the ablated factors, fixing the visual input order is the dominant one: removing the explicit `image -> wrist_image` order   drops LIBERO Goal success rate from **97.2% (486/500)** to **73.8% (369/500)**, i.e. **-23.4 pp**, while keeping the rest of the

setup unchanged.

Experiment	Description	Success / Total	Success Rate	Delta vs Previous
Exp0	Historical 4-in-1 baseline	331 / 500	66.2%	—
Exp1	Fully aligned setup	488 / 500	97.6%	+31.4 pp
Exp2	Exp1 with policy.train() turned off	486 / 500	97.2%	−0.4 pp
Exp3	Exp2 with fixed image order removed	369 / 500	73.8%	−23.4 pp