optimize: improve inference scripts with mmgp and torchao support

[shangvo]

Optimization for Inference Scripts

Changes

• Added inference_mmgp.py for memory optimization using mmgp
• Added inference_torchao.py for model quantization and acceleration
• Updated requirements with new dependencies:
  • para-attn: for first block caching
  • mmgp: for memory management
  • torchao: for model quantization

Benefits

• Reduced memory usage with mmgp profile management
• Accelerated inference with model quantization
• Added support for different hardware configurations

Dependencies

Added new requirements in requirements_new.txt:

• para-attn
• mmgp
• torchao

Please make sure to install these new dependencies before running the optimized inference scripts.