ggml: bump submodule for full CUDA GET_ROWS type coverage

i-quants and mxfp4 join the k-quants on the direct device path, bringing
CUDA to parity with the Metal, Vulkan and SYCL backends.