Topic Relevance: Improve prompt labeling #174

Describe the bug
topic_relevance_llm builds its system prompt by concatenating the tenant's configuration text directly onto the scoring rules (system_prompt = config + "\n\n" + rules in backend/app/core/validators/topic_relevance_llm.py). Nothing tells the model what that text is. With a keyword-style config such as Cricket scores\ncricket\nglific, gpt-4o-mini reads the bare words as noise rather than a forbidden-topic list, so "How do I build a Glific flow?" scores 3 and passes.

This surfaced after #166 made the forbidden schema (v2) the default and kaapi-backend #1245 started sending inline specs, so tenants now hit this path with bare keyword lists and no descriptive stored config.

The validator also breaks on reasoning models (#171 moves the default to gpt-5.6-sol):

• content.strip() raises AttributeError when the completion content is None, reported as LLM call failed.
• The fixed max_tokens=300 cap counts hidden reasoning tokens, so the JSON verdict can be truncated.

Expected behavior

• The config is clearly labelled in the prompt (e.g. inside a <forbidden_topics> / <allowed_topics> block placed after the instructions), and bare keywords are treated as topics.
• None content from the LLM fails gracefully with LLM returned an unparseable response.
• No max_tokens cap; reasoning_effort is sent only when the model supports reasoning.

Steps to reproduce

1. Create a topic_relevance_llm config with prompt_schema_version=2, threshold=2, configuration Cricket scores\ncricket\ncricket match\nglific\nGlific flows.
2. Validate How do I build a Glific flow? with gpt-4o-mini.
3. Observe score 3 / passed; expected score 1 / blocked.

Proposed fix

• Move the config into a {{TOPIC_CONFIGURATION}} placeholder at the end of v1.md, v2.md, v3.md and restructure the prompts per OpenAI's prompting guides (headers, explicit rules, few-shot examples, XML boundary for injected content).
• (content or "").strip() before parsing.
• Drop _MAX_TOKENS; add reasoning_effort="medium" when litellm.supports_reasoning(model) is true.