Refactor diffusion modules and improve training pipeline

[Kim-JeongHan]

Summary by CodeRabbit

• New Features

  • Added Hydra-based configuration system for training.
  • Introduced enhanced checkpoint management with improved loading and saving capabilities.
  • Added device-aware model loading and GPU/CPU synchronization.

• Bug Fixes

  • Fixed device misalignment in neural network inference.
  • Improved state normalization handling.

• Refactoring

  • Migrated internal computation layer from NumPy to PyTorch tensors for improved performance.
  • Consolidated checkpoint resolution and management.
  • Removed legacy command-line training interface.
  • Streamlined configuration structure and removed redundant settings.