Export register_host_buffer/unregister via the backend reg so the existing
GGML_CUDA_REGISTER_HOST path in llama-model-loader pins mmap'd expert
weights. Imports the host pages through VK_EXT_external_memory_host, which
populates device->pinned_memory and lets the existing pinned fast path in
ggml_vk_buffer_write_2d_async DMA straight from system RAM instead of
bouncing through the staging buffer + blocking host memcpy.
Assisted-by: opencode
Lumpiasty
ad3e8a3dc3