fix(deps): support aarch64 Linux

[Kaweees]

Summary

Torch 2.7.0 pulls in triton 3.3.0, which ships no aarch64 wheels, so uv run fails to resolve on Linux aarch64. torch 2.9.0 is the earliest release whose triton dependency provides aarch64 wheels. Lockfile regenerated accordingly.

Linked Work

• Issue:
• Milestone:

Validation

• make check
• uv run pytest -m "not slow"
• Ran on an aarch64 Linux device (NVIDIA DGX Spark)
  Commands actually run:

# paste exact commands here

Impact

• Backend impact: mujoco / motrix / both / none
• Platform impact: macOS / Linux / both / unknown
• Training effect expected: yes / no / unknown

Artifacts

• W&B:
• benchmark result:
• video / screenshot:
• ONNX / checkpoint:

Checklist

• Added or updated tests where needed
• Updated docs if behavior or workflow changed
• Linked the driving issue
• Noted any follow-up work explicitly

[TATP-233]

I agree with the goal of supporting Linux aarch64, but I don’t think this PR should change the project-wide Torch version.

Please keep the default torch==2.7.0 unchanged and scope any newer Torch requirement only to Linux aarch64 if it is strictly needed for wheel
resolution.

In short:

1. Add Linux aarch64 support.
2. Do not change the default Torch version for existing platforms.

[Kaweees]

As mentioned above, UniLab with out of the box support for aarch64 is not feasible with the current version of torch bundled. What workaround do you suggest?