Improving Semantic Control in Discrete Latent Spaces with Transformer Quantized Variational Autoencoders