feat: add paper retrieval tools

[pankajrajdeo]

Agentomics can receive fetched PDFs, but the main agent cannot efficiently search their contents or inspect relevant figures and pages. This adds the downstream retrieval layer for the shared/fetched_papers handoff expected from #163.

Changes

• process PDFs deterministically into cached Markdown and word chunks keyed by content hash
• rank bounded results locally with TF-IDF while preserving source, page, chunk, and processed-file provenance
• return no result when a query has no lexical match
• render only retrieved pages to PNG on demand
• add inspect_image(image_path, query), which returns the image to the same Agentomics model through Pydantic AI multimodal tool content
• restrict image reads to the current run and cap them at 20 MB
• register both tools by default and include the parser/renderer dependencies in the runtime environment

This PR covers text-based PDFs and requires the configured model to accept image input. OCR and page navigation for scanned/image-only PDFs, paper discovery, reader sub-agents, and Q&A orchestration remain follow-up work.

Validation

• focused unit tests: 2 passed, covering retrieval, caching artifacts, blank PDF pages, no-match queries, image return structure, and path restrictions
• live Codex pipeline on a synthetic paper: verified the recorded retrieve_paper_chunks → inspect_image call sequence and the combined text/image answer
• live Codex pipeline on the real Agentomics paper: retrieved the test-isolation evidence, rendered and inspected the page, and correctly explained when the isolated test set is used
• GitHub CPU test suite passed on the initial revision and is rerunning for the review fixes

The branch is based on current main and remains independent of #163 until its fetch implementation is ready to integrate.