Method FProp

Applies SymbolInsertionLayer. We take in a `x`, which represents the groundtruth sequence (i.e., English sequence). We return a sampled rollin (observed) canvas (i.e., random subset of the English sequence), as well as the target (indices) for an insertion-based model (i.e., the tar

(self,
            theta,
            x,
            x_paddings=None,
            eos_id=1,
            force_sample_last_token=True)