Llama.cpp: 6.5x faster prompt processing on Arm with a NEON kernel

A new Arm NEON kernel has been added to llama.cpp that dramatically accelerates prompt processing on Arm CPUs for low-bit quantized models, achieving speedups of up to 6.5x for certain formats like IQ3_M. The optimization addresses a performance gap where Arm CPUs previously relied on a slower fallback path lacking dedicated kernels, while x86 processors had optimized AVX-512 and AVX2 implementations. The improvements vary by quantization format and Arm processor generation, with the largest gains on newer, lower-bit formats and minimal impact on formats that already had Arm optimizations.