llama.cpp Release b11678 Introduces 4 GB CUDA Pinned Host Buffer Limit

Release b11678 of llama.cpp introduces a 4 GB limit on the maximum CUDA pinned host buffer size alongside standard cross-platform build assets. This update incorporates minor bug fixes and routine maintenance changes for the library's various backend implementations. This change helps prevent excessive memory consumption and potential allocation failures when interacting with NVIDIA GPUs during local large language model inference. It ensures more stable memory management across diverse hardware setups. The update specifically addresses issue #30449 by capping the CUDA pinned host buffer size, which is critical for optimizing data transfers between host system RAM and device VRAM. Standard build assets for macOS, Linux, Windows, and Android continue to be provided with multiple hardware backend choices.