better manage types from LLM return calls

[filip-michalsky]

hey stagehand team, seeing a few methods return an any type -- this isn't good practice; esp in like the LLM client where you return instead of just a string for message responses.

[filip-michalsky]

specifically in the lines i posted upthread, the line response.choices[0].message.tool_calls; is coming from the response generated by llmClient.createChatCompletion, not openai specifically, which can cause downstream errors.
My suggestion here would be to have createChatCompletion return a Promise containing the chat output (unless you want to have pricing information or whatever but that can be captured in logging, idk separate conversation). You can also have a callTool method that also returns a string but processes tool calling depending on the provider.
That way, if a client downstream wants to use ollama, perplexity, or any other API provider, they can just plug and play themselves easily if they implement their own llmClient