PAFT: A Parallel Training Paradigm for Effective LLM Fine-Tuning