ggerganov/llama.cpp released b11574 with Hexagon DSP optimizations

Llama.cpp release b11574 introduces performance optimizations for Qualcomm Hexagon DSP tensor copying and concatenation operations. This update focuses on improving short-row handling and reducing DMA latency for edge devices. These optimizations significantly improve on-device AI inference performance and efficiency on Qualcomm hardware. This benefits mobile and edge deployment ecosystems where power consumption and processing speed are critical constraints. The update implements a threaded DMA and HVX gather-row helper, unrolls gather-row loops, and adds a fastpath for f16 data types. It also resolves random hangs in mdev setups by making vtcm_acquire sync state atomic.