feat(benchmark): add Docling validation harness

[kpeez]

This pull request introduces a comprehensive Docling validation benchmark harness under backend/paperchat/benchmarks/docling_validation/. It provides a CLI for running document conversion and chunking benchmarks, data models for fixtures and results, manifest loading utilities, a retrieval scoring module, and an adapter for interfacing with the Docling library. It also adds a small exclusion to the pre-commit config and a module docstring.

Docling validation harness and CLI:

• Added a new CLI (cli.py) to run the Docling validation benchmark, supporting arguments for fixtures, queries, output directory, and top-k metrics.
• Introduced __init__.py and a module-level docstring for the benchmark tooling. [1] [2]

Core logic and integration with Docling:

• Implemented docling_adapter.py, which loads Docling components, runs document conversion and chunking, normalizes chunk data, and handles warnings/errors robustly.

Data models and manifest handling:

• Defined structured dataclasses in models.py for fixture documents, queries, chunking results, summaries, and recommendations, ensuring type safety and clarity.
• Added manifests.py to load and validate fixture and query manifests from JSON files, with error handling for missing or malformed data.

Retrieval and scoring utilities:

• Created retrieval.py to tokenize text, compute term frequencies, cosine similarity, and rank chunks for retrieval evaluation.

Tooling and configuration:

• Updated .pre-commit-config.yaml to exclude large PDF files in test fixtures from pre-commit checks.