Update policy inference script for VR-recorded datasets

[Copilot]

The existing policy script was built for velocity-based (keyboard/spacemouse) datasets. VR teleoperation records a fundamentally different action space — absolute Cartesian pose + gripper — requiring corresponding changes to inference-time action construction and robot commanding.

Action space changes

	Old (velocity)	New (VR/position)
Dims	6 (linear + angular twist)	8 (pos xyz + quat xyzw + gripper)
Control mode	MODE_VELOCITY	MODE_POSITION
Stiffness / Damping	100 / 40	60 / 50 (matches VR teleop)
Gripper	—	/gripper_commands (hande_left_finger_joint)

Changes

• Imports: replaced Twist with Pose; added Wrench (was missing) and JointState
• __init__: creates self.gripper_pub on /gripper_commands
• New method set_vr_cartesian_pose_target: replaces set_cartesian_twist_target; sends an absolute pose via MODE_POSITION with VR-matching impedance params and co-publishes the gripper JointState
• insert_cable: slices the 8-dim unnormalized action — [0:3] position, [3:7] quaternion, [7] gripper — and dispatches via the new method

# Before — velocity twist, 6D
twist = Twist(linear=Vector3(x=action[0], ...), angular=Vector3(x=action[3], ...))
motion_update = self.set_cartesian_twist_target(twist)  # MODE_VELOCITY

# After — absolute pose, 8D
pose.position.{x,y,z}    = action[0:3]
pose.orientation.{x,y,z,w} = action[3:7]
gripper_pos               = action[7]
motion_update = self.set_vr_cartesian_pose_target(pose, gripper_pos)  # MODE_POSITION