feat(pipeline): fused template-gen to streaming dataset command

[tarekio]

Adds pagen pipeline: one command that generates quality- and diversity-gated Arabic document templates via LLM, then streams one fresh fill per image straight into the render pipeline. pagen dataset is left byte-for-byte untouched.

• pagen/diversity.py: diversity metrics promoted out of eval/ so the inline template gate and the offline sweeps run the same code (eval modules become thin re-export shims).
• Template budgeting: plan_template_budget sizes doc-types by breadth from the requested image count, and invents new Arabic doc types via LLM when the built-in pool is exhausted (now spanning academic/scientific and literary genres too).
• Streaming fills: one LLM call per image (no fill reuse), with a per-image debug.jsonl provenance sidecar (font, template, augment path, ground truth).
• Graceful stop: a single Ctrl+C halts cleanly and leaves a valid partial labels.json; a populated output dir is refused unless --append.