* ggml-backend : fix async copy from CPU * cuda : more reliable async copy, fix stream used when the devices are the same