Tokenized Dataset with Sliding Window

A pretraining run is a function from token ids to gradients. This lesson builds the conveyor that feeds the ids in.