Refactor groundmark as thin wrapper on anchorite

[lgruen-vcgs]

Summary

Rewrites groundmark (forked from gemini-ocr) as a batteries-included wrapper around anchorite. Give it a PDF + model string, get annotated Markdown with embedded bounding box coordinates.

What groundmark provides (concrete anchorite provider implementations):

• PydanticAIMarkdownProvider — PDF-to-Markdown via any Pydantic AI model, with a line-number workaround for Claude's content filter
• PdfplumberAnchorProvider — line-level bbox extraction via pdfplumber (MIT), with table detection and ligature decomposition

What anchorite provides (delegated entirely):

• Smith-Waterman alignment, annotation (<span> injection), stripping, fuzzy quote resolution

Key changes from gemini-ocr

• Replace Gemini SDK → Pydantic AI (any LLM via model string)
• Replace Document AI → pdfplumber (local, no API dependency, MIT licensed)
• Replace inline alignment/annotation/resolution code → anchorite
• Delete Airflow scripts, GCS support, settings.py, run_ocr.py
• Delete Sphinx docs (4 source files don't need a docs site)
• Single async process() entry point with chunking support
• Debug visualizer using pypdf Highlight annotations
• Reset version to 0.1.0

What's left (4 source files)

File	Lines	Role
process.py	79	Entry point — wires providers into anchorite.process_document()
markdown.py	94	MarkdownProvider — Pydantic AI + prompt + line-number workaround
parse.py	85	AnchorProvider — pdfplumber extraction with table detection
visualize.py	165	Debug CLI — overlays raw vs aligned boxes on PDF

Test plan

• uv run pytest tests/ -x -q — 6 tests pass
• uv run ruff check src/ tests/ — clean
• Verified visualize.py runs against a real PDF with the anchorite fix branch
• End-to-end test with flowa (next step after this lands)