Implement LLM fallbacks #352
Description
We should implement a fallback system to the LLM .invoke calls. It should abstract the .invoke calls to a wrapper which catches any errors thrown. If an error is thrown, it switches to a different LLM provider, and retries the request. It should keep falling back to different providers until we've exhausted all providers (OpenAI, Google, Anthropic) Don't just switch models, but actually switch entire providers. Ensure this is applied to every LLM call inside the open-swe app (inside the graphs/ directory.
Just pick one model as a fallback for each provider.