Feature: Add Mock HuggingFace Dataset Support for TorchTitan

[Xiaoming-AMD]

Summary

This PR adds mock HuggingFace dataset support for the TorchTitan backend, allowing offline runs
and CI testing without requiring dataset downloads from the HuggingFace Hub.

🔧 Changes

• primus/utils/mock_hf_dataset.py
  • Implements lightweight synthetic dataset generators for token/text samples.
  • Activated automatically when:
    • training.mock_data=True is set in experiment config / CLI.
• primus/modules/trainer/torchtitan/patch_utils.py
  • Provides a wrapper patch_mock_hf_dataset() for backend integration.
• tests/modules/trainer/torchtitan/test_patch_mock_hf_dataset.py
  • Unit tests covering dataset creation, iteration, and shape validation.

🧠 Usage

Real vs Mock dataset toggle via command line:

# Use real dataset (default)
EXP=examples/torchtitan/configs/MI300X/llama3.1_8B-BF16-pretrain.yaml bash examples/run_pretrain.sh --training.mock_data=False

# Use mock dataset (offline)
EXP=examples/torchtitan/configs/MI300X/llama3.1_8B-BF16-pretrain.yaml bash examples/run_pretrain.sh --training.mock_data=True