feat: sync SDK with current worker ZIP contract and agentic retrieval API

[suguanYang]

Summary

Syncs the Python SDK with two server-side changes:

Issue #21 — Align ParseResult ZIP models with current worker output

• DocNav model — parses doc_nav.json from result ZIPs (sections tree + image/table resource summaries)
• Manifest HIERARCHY — exposed via hierarchy field with alias="HIERARCHY"
• document_top_summary — parsed from chunk metadata for all chunk types
• Legacy cleanup — table_type, chunks_slim, and hierarchy marked as legacy; parser handles their absence gracefully
• save() writes doc_nav.json and uses by_alias=True for manifest round-trip

Issue #22 — Add agentic retrieval response fields

• answer_text and referenced_chunks on RetrievalQueryResponse
• use_agentic parameter on both sync and async query()
• All new fields optional for backward compatibility

Test plan

• All 60 existing + new tests pass
• Staging API verified: use_agentic=True returns answer_text (736 chars) + 13 referenced_chunks with chunk_id, document_id, asset_url
• Legacy response (no agentic fields) parses without error
• doc_nav.json round-trips through parse + save
• Parsing succeeds without chunks_slim.json or hierarchy.json

🤖 Generated with Claude Code