fix: preserve full parsed chunks in result ZIP

[suguanYang]

Summary

• keep the full parsed chunk list for result ZIP generation, raw artifact upload, and job_chunks
• compute a separate deduped publication chunk list for retrieval document_chunks and graph publication
• move local media GC until after the result ZIP and raw artifacts are uploaded so duplicate image/table assets are not removed before export
• add a worker contract regression covering DB dedup with duplicate text and image chunks

Closes #28

Tests

• uv run pytest apps/worker/tests/contract/test_parse_task_contract.py::test_should_export_full_result_when_publication_deduplicates_existing_chunks
• uv run pytest apps/worker/tests/contract/test_parse_task_contract.py
• uv run --all-packages --group lint ruff check apps/worker/app/core/tasks/kb_tasks.py packages/shared-python/shared/services/job_lifecycle_sync.py apps/worker/tests/contract/test_parse_task_contract.py
• uv run --all-packages --group typecheck pyright --project pyproject.toml apps/worker/app packages/shared-python/shared