Abstracted Shapes as Tokens -- A Generalizable and Interpretable Model for Time-series Classification