Assistant: handle content-policy refusals with a polite domain redirect instead of "unavailable"

Parent epic: #1288

When the LLM triggers a content warning (e.g. asking how to do gain-of-function research on SARS-CoV-2), the app shows "The analysis assistant is currently unavailable. Please try again later." with a retry. Changing the subject works fine.

There's no content-refusal handling today: a provider content warning surfaces as an exception and is mapped to a generic error (`assistant.py:91-96` → 503/500), which the frontend renders as the catch-all unavailable/retry message (`useAssistantChat.ts:175`).

Instead, detect content-policy refusals and respond with a polite message that the topic is outside BRC Analytics' domain, pointing the user to how to use the assistant — without the "unavailable"/retry framing.

## Acceptance criteria

- A content-policy refusal yields a polite in-conversation message explaining the topic is out of scope and how to use BRC Analytics — not an "unavailable"/retry error.
- Normal conversation continues afterward (the refusal doesn't break the session).