Runs ONNX models on any Nx backend.
Moves a model's params to EMLX and keeps 4-bit and 8-bit weights packed.