Add a Cohere LLM adapter

[jstuart0]

Goal

Ship a Cohere adapter (command-r, command-r-plus) for users who prefer Cohere's RAG-friendly models.

Background

Same shape as issue #1 (Mistral). Cohere has its own Chat API that's JSON-ish but structurally different from OpenAI's chat completions — it uses `{"message": "...", "chat_history": [...]}` and the response shape is `{"text": "..."}` instead of a choices array.

Where to start

• `src/server/services/ai/llm/cohere.ts` — implement `createCohereAdapter({ apiKey })`.
• Translate our `LlmRequest.transcriptPrompt` into a Cohere message + systemPrompt → `preamble`.
• Add `"cohere"` to `ProviderKind` + pricing table.

Acceptance criteria

• Configured Cohere provider passes a watcher run end-to-end.
• Handles Cohere's error shape (`{"message": "error body"}`).
• Tests mock the `/chat` endpoint covering happy-path + non-200.

Extra credit: implement `completeStream()` (Cohere streams with SSE-ish `event: text-generation` deltas) so the Ask assistant benefits from it.