Align JEPA supervision with leakage-free future targets

[mehulsuresh]

Summary

This PR adds an opt-in JEPA supervision path for lerobot_datasets that constructs future-shifted target clips and encodes them separately from the context clip.

When datasets.vla_data.video_target_shift_steps is 0, behavior is unchanged.

When video_target_shift_steps is set to one encoder tubelet, the dataloader emits:

• a shorter context clip for the predictor input
• a future-shifted target clip with the same encoded temporal length for supervision

VLA_JEPA.forward() then encodes those clips separately instead of deriving both context and target states from a single encoded video window.

Paper Alignment

The paper describes VLA-JEPA as "leakage-free state prediction" and says future frames should be used only as supervision targets, not as inputs to the learner (paper, HTML).

This PR moves the implementation toward that setup:

• the predictor input comes from a context clip
• the supervision target comes from a separate future-shifted clip
• the default single-pass path remains unchanged unless video_target_shift_steps is enabled