feat: add GLM-OCR SDK service adapter for local MLX deployment

[jonathan-teamstatus]

Summary

• Add a lazy public service config/factory and structured GLM-OCR SDK adapter (not a raw chat endpoint).
• Preserve native labels/reading order, map normalized geometry to source pixels, validate malformed output, and state cancellation/token-limit boundaries.
• Add mocked tests and a sanitized real synthetic-page response. No ML runtime dependency or version bump.

Validation

• 38 upstream tests pass; focused Pyright/pylint pass.
• Real Apple M4 Pro: Metal MLX inference + CPU layout + localhost SDK; non-loopback Python network guard enabled.
• Paired PDF Craft integration produces Markdown/EPUB and reusable PCEX with services stopped.

Limitations / release gate

• SDK defaults discard footnotes; explicit retained-label config and disabled formatter merges are required.
• Layout missed a synthetic equation; rendered footnote fidelity not validated (model altered/missed markers).
• SDK usage is empty; token budgets are rejected, not estimated.
• Draft pending maintainer review and upstream release; do not merge automatically.