feat: Implement gradient difference unlearning support

- Added gd_mode and gradient difference parameters to neox_args
- Modified training.py to support gradient difference in interleaved mode
- Updated data_utils.py to load retain dataset for gradient difference
- Modified eval.py for PyTorch weights_only parameter compatibility
- Updated test_gradient_ascent.py to include gradient difference tests
- Cleaned up temporary analysis files and scripts

Current implementation uses gd_retain_weight as forget weight multiplier
(L_total = L_retain - α * L_forget), which will be inverted next

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <[email protected]>

Author: Kyle1668