data ingest: wire the CLI-pending tasks — 5 text layouts + semantic_segmentation (RFC-0002 phase 4)

[LukasWodka]

Phase 4 of RFC-0002. Wire the CLI-pending tasks so the CLI matches the platform (schema + ingestor + backend already support them).

• Text family: token_classification, sentence_pair_classification, causal_language_modeling, seq2seq, embeddings — their texts/ discover + staging + the task-aware label handling (self-supervised = no label).
• semantic_segmentation — its blocker (data-ingestors#136) is closed; add masks/ sidecar staging + the mask_id column (mirror object_detection's annotations path). Note the training-side sign-off is tracked in backend#816.

Design-of-record: RFC-0002 (#174). Part of the data-ingest redesign epic. (§7, §11 phase 4)

[LukasWodka]

Re-scoped per RFC-0002 Principle 6 + Option A (cli#174 §14): the per-task layout rules are the ingestor's, not the CLI's. They move to a machine-readable layout contract in data-ingestors#347. This ticket becomes CLI-side only: consume that contract to discover + stage the pending tasks (interface + contract-locked preview), rather than hardcoding 6 layouts in Go. Depends on data-ingestors#347.