Closing the Performance Gap in Generative Recommenders with Collaborative Tokenization and Efficient Modeling