chore(db): remove deprecated entity schema from memory_links

[nicoloboschi]

Summary

Closes #3005.

Entity edges are no longer materialized in memory_links. Retain stores memory-to-entity associations in unit_entities, and both read paths derive entity edges from that table on demand — the /graph endpoint from shared unit_entities rows, and recall via the unit_entities self-join. Migration e9b2c7d1f3a4 deleted the stored entity rows, and current writers only ever pass entity_id = NULL, so the entity-specific schema on memory_links was dead weight that made the table look like it still supported a second, materialized source of truth for entity edges.

This removes that schema and the write-path plumbing behind it.

Schema changes (new migration, PG + Oracle)

• Defensively delete any residual link_type = 'entity' rows.
• Drop the entity_id column and its FK to entities.
• Drop the entity index (idx_memory_links_entity / idx_ml_entity).
• Recreate the link_type CHECK without 'entity' (temporal, semantic, and the legacy causal values remain valid).
• Collapse the function-based idx_memory_links_unique to (from_unit_id, to_unit_id, link_type) — non-entity rows already deduplicate on those three columns via the old COALESCE(entity_id, nil) key, so effective uniqueness is preserved.

Production-safe by construction

memory_links and its entity index can be very large on established banks, so the migration avoids long exclusive locks:

• the residual delete is chunked with per-batch commits;
• every index is swapped CONCURRENTLY (PG) / ONLINE (Oracle), building the new unique index under a temporary name and renaming so duplicate protection is never dropped;
• the new CHECK is added NOT VALID then VALIDATEd separately, so writers are never blocked on a full-table scan.

Downgrade restores the former schema shape but, as documented in the migration, cannot reconstruct the historical entity rows e9b2c7d1f3a4 already deleted.

Application code

• Remove MemoryLink.entity_id / MemoryLink.entity and Entity.memory_links; update the docstring and CHECK metadata.
• Simplify internal link tuples from (from, to, link_type, weight, entity_id) to (from, to, link_type, weight) across the temporal, semantic, causal, and graph-maintenance producers.
• Remove _NIL_ENTITY_UUID and the nil_entity_uuid DataAccessOps parameter; drop the entity_id column/placeholder from the PG and Oracle bulk inserts (conflict target is now (from_unit_id, to_unit_id, link_type)).
• Drop the dead entity_id tiebreaker from the chunk-storage lock ordering and a stale runtime comment.
• The graph API keeps returning dynamically derived entity edges from unit_entities — unchanged.

Tests

• New test_migration_drop_memory_links_entity.py (dedicated pg0 instance): asserts the column/index/entity CHECK value are gone at head, the unique index is three columns, and the downgrade → re-upgrade round-trip restores and re-drops the shape.
• Updated test_link_utils.py and test_document_transfer.py for four-element tuples / no entity_id column.

Non-goals

Unchanged: the unit_entities model, link_type = 'entity' in graph/API response types, entity resolution / co-occurrence / orphan cleanup, and temporal/semantic/causal link generation.