llama.cpp b9572 fixes rms_norm backprop with in-place aliasing

The latest release of llama.cpp, tagged b9572, addresses a critical bug in the rms_norm_back function—the backward pass of root mean square normalization. The error occurred when in-place aliasing of tensors caused the gradient computation to produce incorrect output, potentially corrupting model updates during training or fine-tuning. The fix, contributed by Georgi Gerganov, includes a cleaned-up comment for clarity. This is a significant quality-of-life improvement for developers using llama.cpp for local LLM inference and training, as rms normalization is widely used in modern architectures (e.g., Llama, Mistral).

b9572 is packaged for nearly every major platform: macOS (Apple Silicon, Intel, iOS), Linux (x64, arm64, s390x with CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32), Windows (x64 CPU/ARM64, CUDA 12/13, Vulkan, HIP), Android arm64, and even openEuler. This broad support underscores llama.cpp's role as the go-to inference engine for running LLMs on commodity hardware. For professionals deploying local models, this patch ensures more reliable training loops and consistent output when leveraging advanced features like in-place tensor operations.