llama.cpp b11059 optimizes Metal FWHT for Apple Silicon

llama.cpp has released version b11059, focusing on performance optimizations for the Metal backend on Apple Silicon hardware. The update modifies the Fast Walsh-Hadamard Transform (FWHT) kernel to accept F16 inputs natively, eliminating costly data type conversions and reducing latency by approximately 3% in benchmarked Hadamard matrix multiplications. Additionally, the release refactors dispatch predicates into shared common headers to resolve compilation errors on macOS ARM64 and ensures robust handling of contiguous tensor shapes. These changes improve inference efficiency for local LLM users running models with Hadamard-based operations on M-series Macs.