Fine-tuned motions cannot be replayed: no motion_lib to deploy reference export step

[Harshita-82]

Summary

There is no documented path from a training motion_lib motion to the deploy
reference format, so a motion a policy was fine tuned on cannot be replayed in
sim2sim or on hardware.

1. No export step from motion_lib to deploy reference format

gear_sonic_deploy/reference/convert_motions.py consumes a PKL keyed by:

joint_pos       (T, 29)      joint_vel       (T, 29)
body_pos_w      (T, 14, 3)   body_quat_w     (T, 14, 4)
body_lin_vel_w  (T, 14, 3)   body_ang_vel_w  (T, 14, 3)

The PKLs produced by the documented data pipeline
(gear_sonic/data_process/convert_soma_csv_to_motion_lib.py) instead contain
root_trans_offset, pose_aa, dof, root_rot, smpl_joints, fps.
There is no key overlap, and no shipped script bridges them.

To be precise about where the gap is: MotionLibBase already computes every
array the deploy format needs. fk_batch resamples to target_fps via
interploate_pose and derives velocities, and motion_lib_base.py (~L1599-1619)
converts quaternions xyzw -> wxyz and slices to the selected body subset at load
time. The arrays exist in memory during training and are used for tracking
rewards. The only missing piece is a step that writes them to disk.

deploy.sh defaults to MOTION_DATA_DEFAULT="reference/example/", so when no
converted motion exists the stack loads the bundled examples and emits no
warning:

Motion Data:  reference/example/
Found 13 motion folders
✓ Loaded dance_in_da_party_001__A464_M (497 timesteps)
✓ Loaded macarena_001__A545 (1375 timesteps)
...

Proposed fix

PR #239 adds gear_sonic/data_process/export_motion_to_deploy.py, which loads a
motion through MotionLibRobot, reads the arrays the library already computes,
and reuses convert_motions.py for CSV writing. It deliberately does not
reimplement forward kinematics, resampling or differentiation.

Verified against MuJoCo's independent forward kinematics across 7 motions
(~3900 frames): exported body poses agree to 2.3e-06 m and 2.1e-04 deg,
which is the float32 / 6-decimal-CSV noise floor. The exported motions were also
replayed end-to-end in sim2sim with a fine-tuned checkpoint and track correctly.

If there is an intended path for this that I have missed, I am happy to close
this out — I could not find one in the docs or scripts.

Environment

Repo        main @ aa263a8
OS          Ubuntu 24.04.4, kernel 6.17.0-35
GPU         RTX PRO 6000 Blackwell Server Edition, driver 595.58.03
CUDA        12.6
Isaac Sim   5.1.0.0
Isaac Lab   2.3.0
Python      3.11.15
PyTorch     2.7.0+cu128
MuJoCo      3.11.0