[codex] Add action-labeled bilinear ablation tooling

[fei-yang-wu]

Summary

• Add action-labeled Dance102 bilinear ablation tooling, including the labeled manifest, cluster sweep wrappers, policy-label export, offline-pretrain evaluation, and deterministic rollout recording helpers.
• Add latent goal observation/task support plus VQ-VAE and goal-AE bilinear config variants for the G1 latent tasks.
• Sync submodule pointers to pushed commits: ImitationLearningTools 2aa832b for NPZ action-label loading and RLOpt 29f749d for bilinear latent pretraining/BC controls and latent learner checkpointing.

Submodule PRs

• [codex] Add NPZ action labels to LAFAN loader ImitationLearningTools#2
• [codex] Improve bilinear latent pretraining controls fei-yang-wu/RLOpt#15

Validation

• git diff --check in the parent repo
• git diff --check in ImitationLearningTools and RLOpt
• bash -n on updated cluster/experiment shell scripts
• python3 -m py_compile on the new RLOpt helper scripts

Notes

Generated NPZ label/rollout/debug outputs are intentionally not tracked. The action-label manifest remains as a small checked-in recipe/path contract, and the docs now describe the label NPZ as a local/generated artifact.