Masked Autoencoding Does Not Help Natural Language Supervision at Scale