Re-scan legacy articles with the embed_detector (added in commit da3d16d) and populate parsed_json['embeds'] for rows ingested before the detector existed.

Body:

• since: optional ISO datetime — only consider rows created after this (default: 60 days ago).
• limit: max rows to scan (default 200, max 5000).
• dry_run: if True (default), count only — no writes.
• only_if_no_embeds: if True (default), skip rows whose parsed_json['embeds'] is already a non-empty list.

Returns counts and a by_type breakdown of newly detected embeds. Per-row exceptions are caught + logged — one malformed article never poisons the batch.