Template detection

[gavinbee]

`src/template_detect.py` — classify page 1 of the input PDF against the template registry.

• Rasterize page 1 (`pdf_io`), send to `qwen2.5vl:*` with a classification prompt: "Which of these On-Deck Evaluation templates is this? Respond with template_id and confidence." Enum comes from `TEMPLATES` keys.
• If confidence < threshold (default 0.7) or model says `unknown` → exit with a clear message offering `--template ` override and asking the user to file an issue with the PDF.
• `--template ` short-circuits detection.

Tests use a mocked vision client.

Depends on #6, #8.