rank-1-deflation

Introduces AdaPaD, a parallel rank-1 deflation method for LoRA fine-tuning, enabling low-rank linear regression components to be computed concurrently instead of sequentially, improving efficiency.