llama.cpp b11043 adds HMX flash-attention for non-standard head dims

llama.cpp has released version b11043, introducing support for HMX flash-attention with non-standard head dimensions such as DK=DV=72. This change enables efficient inference for models like SigLIP by padding tail lanes rather than rejecting incompatible shapes. The update also upgrades CUDA library dependencies to 12.8 and 13.3 on Linux and Windows builds, alongside maintaining ROCm 10.0 and OpenVINO support. These changes broaden the range of multimodal and specialized models compatible with local inference without sacrificing performance.