feat: add GitHub Copilot as LLM provider

[CodingIsBliss]

Summary

Enable using a GitHub Copilot subscription to power MiroFish's LLM calls — no separate API key needed.

Motivation

Many developers already have a GitHub Copilot subscription. This PR lets them use that same subscription to run MiroFish simulations, removing the need to set up and pay for a separate LLM API provider.

This uses the same token exchange mechanism as OpenClaw's built-in github-copilot provider.

How it works

1. Exchanges a GitHub token for a short-lived Copilot API token via api.github.com/copilot_internal/v2/token
2. Derives the OpenAI-compatible base URL from the token's proxy-ep field
3. Caches tokens in memory + on disk, auto-refreshes before expiry
4. Thread-safe for concurrent simulation agent calls
5. Zero changes needed for existing users — only activates when LLM_PROVIDER=github-copilot is set (or when no LLM_API_KEY exists but a GitHub token is detected)

Configuration

LLM_PROVIDER=github-copilot
GITHUB_TOKEN=ghp_xxx
LLM_MODEL_NAME=gpt-4o

Also supports GH_TOKEN and COPILOT_GITHUB_TOKEN env vars (priority: COPILOT_GITHUB_TOKEN > GH_TOKEN > GITHUB_TOKEN).

Files changed

File	Change
backend/app/utils/copilot_auth.py	NEW — Token exchange, caching, auto-refresh
backend/app/utils/llm_client.py	Copilot auto-detection + token refresh before each call
backend/app/config.py	Added LLM_PROVIDER config key
backend/app/utils/__init__.py	Export new module
.env.example	Copilot setup instructions

Notes

• Uses only stdlib urllib for the token exchange (no new dependencies)
• Copilot models available depend on the user's plan (Free/Pro/Business/Enterprise)
• Rate limits are lower than dedicated LLM APIs — users should start with small simulations (<40 rounds)
• The copilot_internal API is undocumented but stable (used by VS Code Copilot + OpenClaw)