sync : ggml

vulkan: add noncontiguous GLU support (llama/21081)

hexagon: support for IQ4_NL and MXFP4 (llama/21018)

rpc : proper handling of data pointers to CPU buffers (llama/21030)

metal : Fix dimension constraint violation in matmul2d descriptor (llama/21048)

hip: use fnuz fp8 for conversion on CDNA3 (llama/21040)

opencl: allow large buffer for adreno (llama/20997)

fix(ggml): correct RISC-V ISA string canonical ordering for RVV in CMake (llama/20888)

ggml-cuda: Add NVFP4 dp4a kernel (llama/20644)

CUDA & CPU: support F32 kernel type for `CONV_TRANSPOSE_2D` (llama/17094)

mtmd: Add DeepSeekOCR Support (llama/17400)

llama: fix llama-model-saver (llama/20503)

sycl : fix wrong variable check by assert (llama/20903)

metal : add FLOOR, CEIL, ROUND, TRUNC unary ops (llama/20930)

metal : add FA instantiations for HSK=512, HSV=512 (llama/20902)