Remove ~6.5MB of redundant Word data from Hatchet task boundaries:
- Remove words from TopicChunkInput/TopicChunkResult (child workflow I/O)
- detect_topics maps words from local chunks by chunk_index instead
- TopicsResult carries empty transcript words (persisted to DB already)
- extract_subjects refetches topics from DB instead of task output
- Clear topics at detect_topics start for retry idempotency
