Fix documented action layout

[sawhney17]

What

Correct the documented seven-value action layout from six joint velocities plus gripper position to six commanded Cartesian-position values plus commanded gripper position.

Evidence

The pinned dataset-builder implementation constructs action with:

np.concatenate((action['cartesian_position'], [action['gripper_position']]))

See droid.py lines 69–77. The same builder's feature documentation currently carries the joint-velocity wording, so the generated features.json inherited it.

I also decoded all 100 episodes in the official droid_100/1.0.0 sample and compared every step. All 32,212/32,212 action vectors exactly equal the six-value action_dict.cartesian_position concatenated with the one-value action_dict.gripper_position; the maximum absolute difference is 0. action differs from action_dict.joint_velocity (maximum absolute difference 4.125915944576263).

The machine-readable audit, pinned object generations/checksums, and RFC 6902 metadata patch are published in droid-quality-audit v0.1.0:

• all-step action/reward audit
• description patch
• source-object lock

Scope

This PR changes one documentation line only. It does not alter stored dataset values or claim that the action is Cartesian velocity.