Bilevel optimization for data-driven learning of Koopman embeddings using kernel-based autoencoders