Add seq2seq / text-generation task family #285

Add a seq2seq / text-generation task family, with BLEU (token-sequence metric) as the named POC consumer that proves the new output structure flows end-to-end.

Context

Task families live in src/raitap/task_families/ (base.py, registry.py); existing: classification, detection. A task family defines how model outputs are interpreted, how predictions/targets are formed, and which metrics/visualisers apply. seq2seq (encoder-decoder / autoregressive generation) is a distinct output structure (token sequences, not class logits).

Scope

• A seq2seq task family: output = generated token sequence(s); define prediction extraction, reference/target handling.
• Wire it into the task-family registry; capability/compat gating for seq2seq.
• Decide the supported explanation surface (token-level attribution links to Add batch-aware text attribution visualisation #99 text attribution visualisation) and which explainers/assessors are compatible.

POC consumer

• BLEU (or exact-match as the minimal variant) is the concrete POC metric: a small seq2seq model generates sequences and BLEU is computed against references end-to-end. This proves the token-sequence output structure threads through prediction extraction -> metric. Add batch-aware text attribution visualisation #99 token-level attribution is the deeper transparency consumer, follow-on.

Acceptance

• A small seq2seq model runs through the pipeline producing generations + a BLEU score end-to-end.
• Compat gates correctly include/exclude methods for the seq2seq output shape.
• Tests + docs.

Relationships

• Add batch-aware text attribution visualisation #99 (batch-aware text attribution visualisation) is the natural transparency consumer, follow-on.
• Likely needs the text input modality / tokenisation path; coordinate with existing text-modality work.