feat: Add SONIC locomotion controller (CPU inference via onnxruntime)

[MiaoDX]

Context

Follow-up from #82. The SONIC controller from NVlabs has been investigated and deferred due to architectural complexity.

Research Findings

SONIC ONNX models are available at nvidia/GEAR-SONIC on HuggingFace. The C++ inference code in NVlabs/GR00T-WholeBodyControl confirms CPU inference is feasible — OrtSession uses CPUExecutionProvider with no GPU/CUDA references.

However, the architecture is significantly more complex than GR00T/Holosoma:

Aspect	GR00T / Holosoma	SONIC
ONNX models	1-2 single models	3 models (encoder + decoder + planner)
Input tensors	1 (obs vector)	6-11 (context qpos, velocity, mode, direction, seed, etc.)
Observation	Single vector (86-100D)	4-frame context window (4 × 36) + command inputs
Locomotion modes	Balance/Walk or single policy	IDLE, SLOW_WALK, WALK, RUN, BOXING

Proposed Implementation

Add SonicLocomotionController to src/roboharness/controllers/locomotion.py:

1. Planner-only mode: Use planner_sonic.onnx for locomotion (keyboard/gamepad-style velocity commands → joint targets). This is the simplest entry point and most analogous to GR00T/Holosoma.
2. Full pipeline mode (stretch): Wire encoder + decoder for motion tracking / VR teleoperation inputs.

Key details for planner (LocalMotionPlannerONNX):

• Input (v0, 6 tensors): context_mujoco_qpos [4×36], target_vel [1], mode [1], movement_direction [3], facing_direction [3], random_seed [1]
• Output: Predicted qpos frames [N×36] + frame count
• Modes: 0=IDLE, 1=SLOW_WALK, 2=WALK, 3=RUN, 4=BOXING

References

• HuggingFace: https://huggingface.co/nvidia/GEAR-SONIC
• GitHub: https://github.com/NVlabs/GR00T-WholeBodyControl/tree/main/gear_sonic_deploy
• Related: CI Strategy: Add GPU testing support for cuRobo / Policy / WBC #18 (GPU-dependent work)

Parent issue: #82