Fix resume bug

[Choiyoonji]

Fixes #229

Changes proposed in this pull request:

• This PR addresses an _pickle.UnpicklingError that occurs when resuming training using weights_only=True with checkpoints saved in PyTorch 2.1+.

• Specifically, the following code was added right after import torch and import numpy as np in /gr00t/experiment/trainer.py:

  torch.serialization.add_safe_globals([
      np.core.multiarray._reconstruct,
      np.ndarray,
      np.dtype,
      np.dtypes.UInt32DType
  ])

• This change allows PyTorch to safely unpickle RNG state and resume training by explicitly allowlisting required NumPy globals.

• Example error message:

  _pickle.UnpicklingError: Weights only load failed...
  WeightsUnpickler error: Unsupported global: GLOBAL numpy.dtypes.UInt32DType was not an allowed global by default.
  

---

Error Reproduction & Test Case

• How to reproduce:

  1. Start training and interrupt it manually (e.g., using Ctrl+C) to generate a checkpoint.

  2. Attempt to resume training using the --resume flag.

  3. You will encounter the following error:

     _pickle.UnpicklingError: Weights only load failed...
     

• Fix validation:

  • After applying the changes in this PR, resuming training using the same checkpoint works correctly without errors.
  • If similar errors occur for other NumPy types (e.g., Float64DType, Int64DType), they can be added to the allowlist in the same way.

---

Reference

• [PyTorch Documentation – torch.load (Security Section)](https://pytorch.org/docs/stable/generated/torch.load.html#security)

---

Before submitting

• I've read and followed all steps in the Making a pull request
  section of the CONTRIBUTING docs.
• I've updated or added any relevant docstrings.
• If this PR fixes a bug, I've added a test that will fail without my fix.
• If this PR adds a new feature, I've added tests that sufficiently cover my new functionality.

[youliangtan]

@Choiyoonji Thanks for the contribution!