Delta-CoMe: Training-Free Delta-Compression with Mixed-Precision for Large Language Models