Learning hard quantum distributions with variational autoencoders