llama.cpp Release b10970 Adds FP32 Accumulation for AMD CDNA GPUs

llama.cpp release b10970 updates the HIP backend to use 32-bit floating-point (FP32) accumulators in fused attention matrix-multiply-accumulate (fattn-mma) kernels on AMD CDNA hardware. This update enhances numerical precision and computational stability during LLM attention calculations on AMD Instinct enterprise GPUs. It helps prevent precision loss or overflows when running large language model inference on AMD data center hardware. The patch specifically targets MFMA (Matrix Fused-Multiply Add) instruction execution within the fattn-mma implementation for CDNA architectures (#28576). Release artifacts include updated pre-built binaries across multiple platforms including Linux CUDA, ROCm, Vulkan, and SYCL.