Add offline scene tag inference

[hualuoz]

Summary

• add a deterministic offline scene tagger for background descriptions and image filenames
• optionally infer and merge scene tags in poseguide guide recommend
• cover description, image-fixture, CLI integration, and no-match behavior
• document the offline behavior and a future CLIP upgrade seam

Why

PoseGuide previously required a scene JSON file or manually supplied tags. It had no path from a background image name or natural-language description to the tags consumed by the ranker.

User impact

Users can now request recommendations with --description or --background. The implementation makes no network requests, reads no image pixels, and downloads no model; explicit --tags can still be combined with inferred tags.

Validation

• pytest -q — 46 passed, 5 skipped
• pytest -q --cov=src --cov-fail-under=30 — 70.73% coverage
• ruff check src tests
• ruff format --check src tests
• python -m compileall -q src tests

Fixes #6