refactor(cook): data-driven recipe generation with in-process flatten and lint

Replace the hardcoded TEMPLATES dict and the ksflatten-relative wrapper
with a single data-driven pipeline:

- recipe_templates.yaml becomes the single source of truth with three
  sections: base (always included), choices (exactly-one per category,
  the invariant pykickstart cannot check) and features (additive flags
  or one-of values)
- generate_recipe.py is a pure YAML consumer: it resolves ingredients,
  renders dishes via pykickstart in-process (no ksflatten binary, no
  %include path munging), lints the invariants pykickstart cannot check
  (exactly-one per choice category, known keys, existing fragments,
  unique filenames) and validates every generated dish
- dish names are derived by a generic rule with canonical category order
  from the templates, independent of YAML key order; no committed
  generated files means no name burn-in
- recipes/ and dishes/ become gitignored build products; the 68 tracked
  generated files (already out of sync with the manifest) are removed
- delete dead ingredients (section-data/, validation/, initial-setup/
  desktop/, live/hypervisor.ks, rpmfusion-nonfree.ks), the stale
  all-ingredients.cfg (replaced by 'make inventory') and the
  bin/generate-recipe and bin/ksflatten-relative wrappers
- Makefile: single 'make all' step plus lint, validate, inventory, test
- add 16 pytest tests covering expansion, selection, naming, lint and
  flatten/validate round-trips

.gitignore

@@ -0,0 +1,8 @@
+# Generated build products (created by `make all` in cook/)
+cook/recipes/*.cfg
+cook/dishes/*.cfg
+
+# Python
+__pycache__/
+*.py[cod]
+.pytest_cache/