Optimizing AutoML for the tinyML Future