FASTer: Toward Efficient Autoregressive Vision Language Action Modeling via Neural Action Tokenization