feat: Add AI recipe generator chat

[vanalmsick]

Hi @TomBursch, 👋

Thanks for all the work on KitchenOwl. I use it daily, but coming up with new recipes is always a pain, so I built this AI recipe generator chat feature:

Describe what you're in the mood for (or what's in your fridge), and get a complete recipe - title, times, servings, ingredients, and steps - in your household's language, using items you already have. Refine with quick replies or free text, then save via the existing add-recipe form.

What's new

For users

• Recipes + button gets a third option, ✨ AI (shown only when enabled server-side).
• Chat starts with suggested prompts; each reply is a recipe preview with quick actions (different recipe, make it vegetarian, more detail, or free-text edits).
• Add recipe pre-fills the standard form (same flow as URL import).
• Token usage shown per answer and per session; Stop cancels a slow request.

For self-hosters

• Reuses the existing LLM config (no new dependencies). New env var: LLM_RECIPE_GENERATION (default true if LLM_MODEL is set) to disable the feature.

How it works

• New endpoint POST /api/household/<id>/recipe/generate (JWT + household-scoped, stateless, only user/assistant roles accepted, capped at 50 messages × 20k chars).
• recipe_generation.py calls LiteLLM with a JSON-schema response, falling back to extracting JSON if a provider ignores the schema. Prompt includes household language and existing item names.
• Ingredients matched to existing items by exact, case-insensitive name; unmatched ones become new items on save; duplicates dropped.
• /api/health exposes llm_recipe_generation so the app knows whether to show the button.

Known limitations

• Stop only stops the UI from waiting — the backend call still runs and is billed.
• Full item list is sent every time; fine for now, may need filtering for very large households.
• Token counts shown, but not cost in currency.

Testing

• New backend test covering flag, matching, dedup, usage, and role rejection.
• Full backend pytest passes (one unrelated timezone-based failure).
• flutter analyze clean; UI renders at phone/desktop widths.