LoRA and DoRA Implementation

Parameter-efficient fine-tuning from first principles — every matrix decomposition derived and implemented in PyTorch without libraries. Validated against Hugging Face PEFT outputs for correctness.