Fix ARM64/aarch64 Docker builds for reconstruction and robotic_grounding

[kabilankb]

Summary

• Docker Hub's pytorch/pytorch base images ship amd64-only, and several pinned pip packages (usd-core, judo-rai, onnxruntime-gpu==1.24.4) have no linux_aarch64 wheel, so reconstruction's and robotic_grounding's Docker builds fail outright on ARM64 hosts (e.g. NVIDIA GB10 Grace-Blackwell).
• Retarget v2d_moge and v2d_task_library_loader from pytorch/pytorch:* to NVIDIA's arm64-native NGC PyTorch image, with a numpy<2.0 pin to match that image's torch build ABI, and add the missing libx11-dev system dep (needed by glcontext, which has no arm64 wheel and must build from source).
• Add robotic_grounding/workflow/Dockerfile.aarch64 (wired up by the existing build-aarch64 path in workflow/run.sh, which previously had no corresponding Dockerfile) on top of the already arm64-native Isaac Lab NGC base:
  • Add liblapack-dev/libblas-dev (needed by py-soma-x's cholespy dependency, which has no arm64 wheel).
  • Bump onnxruntime-gpu to 1.29.0 (no arm64 wheel exists for 1.24.4).
  • Drop the judo-rai pip install — it has no arm64 wheel for any version with a source fallback, and is only used as an optional import (try/except ImportError) in the SOMA/G1 whole-body visualizer, not on the Sharpa floating-hand path this image targets.
  • Point PYTHONPATH/LD_LIBRARY_PATH at Isaac Sim's bundled omni.usd.libs extension instead of pip-installing usd-core, which has no linux_aarch64 wheel or sdist on PyPI at all.

Test plan

Verified end-to-end on an NVIDIA GB10 (Grace-Blackwell, aarch64) host:

• v2d_moge image builds and the documented MoGe quickstart (run_download_weights + run_video_to_depth on the bundled test_video.mp4) produces correct depth maps + intrinsics for all 100 frames.
• v2d_task_library_loader image builds successfully.
• robotic_grounding's aarch64 retarget image builds successfully; dummy_agent.py and a 3-iteration train.py RSL-RL PPO smoke test both run correctly against the bundled, license-free synthbox fixture, producing real reward/loss curves and checkpoints.
• No changes to amd64 behavior — all fixes are scoped to the aarch64 path (new Dockerfile.aarch64) or additive (numpy pin, extra system packages) on images that were already broken on ARM64.