onnxruntime

Do not inline ExternOp's scalar tensor inputs (#3426)

An ExternOp's input needs buffers, so we cannot add compute_inline
schedule on it even if it's a scalar tensor. Instead, we need to
schedule it as compute_root.