feat: introduce BaseExternalParser protocol for pluggable OCR/VLM backends

[gavin913-lss]

Problem

Every new OCR/VLM parser engine copies the same four-file structure and adds another if engine == … branch in the pipeline. No shared contract exists between engines.

Fixes #3197

Solution

1. BaseExternalParser protocol (_base.py)

Abstract base class with three methods:

• is_bundle_valid(raw_dir, source_path) — cache-hit check
• download_into(raw_dir, source_path) — fetch raw bundle from external service
• build_ir(raw_dir, document_name) — convert raw bundle to IRDoc

2. Engine registry (_registry.py)

Dict mapping engine names to parser classes:

• register_parser("mineru", MinerUParser)
• register_parser("docling", DoclingParser)

New engines register here instead of adding pipeline branches.

3. MinerU adapter (mineru/adapter.py)

Wraps existing MinerURawClient + MinerUIRBuilder + is_bundle_valid behind the protocol.

4. Docling adapter (docling/adapter.py)

Wraps existing DoclingRawClient + DoclingIRBuilder + is_bundle_valid behind the protocol.

What this does NOT change

• No behavior change for existing MinerU/Docling usage
• No changes to on-disk cache format or IRDoc schema
• No changes to pipeline.py dispatch logic (that's a follow-up)
• The existing parse_mineru/parse_docling methods continue to work

Next steps (follow-up PRs)

1. Add a generic parse_external method to the pipeline that uses the registry
2. New engines (PaddleOCR-VL, DeepSeek-OCR, etc.) implement the protocol
3. Eventually remove the if engine == … chain from the pipeline