Automatic1111 for Apple metal, 40% speed up sd1.5

A developer optimized Automatic1111's performance on Apple Silicon by implementing selective Metal Flash Attention and reducing command buffer submissions, achieving roughly 2-3x speedup (5-7 seconds vs. 8-10 seconds on M3 Pro) while maintaining full compatibility with the original software's features and workflows. The key insight was measuring actual workload performance rather than applying optimizations in isolation, and using custom Metal implementations only where testing proved they provided real benefits.