Major refactoring of the Metal backend to make the implementation safer and more manageable:

• Add ggml-metal-device.h and move device code there
• Add ggml-metal-context.h and move backend context code there
• Migrate most kernels to inference-time compilation
• Split ggml_metal_encode_node into per-op functions and move them to separate sources
• Apply typedef struct ggml_mem_ranges * ggml_mem_ranges_t;