#651 ended up not working, but simply add a Suggestor type that asks the LLM to return a map of misspelled word to correct word. Expensive and only works in the suggestion flow, but better than nothing!
