From CUDA to MLX: How K-Search Brings Kernel Expertise to Apple Silicon