fix: incomplete response handling

[pkgodara]

Runner classified every non-tool-call LLM turn as :final_answer regardless of the response's finish_reason. When a provider returns a streaming response with finish_reason: :incomplete/:error/:cancelled and no text content, the runner accepted it as a successful completion, emitted :request_completed with result: "", and ask_sync/3 returned {:ok, ""}. Instead, it should return error tuple.

I found this issue while building the Greeter from docs.

This is what I defined -

defmodule GreeterAi do
  use Jido.AI.Agent,
    name: "greeter",
    description: "Generates a friendly greeting",
    tools: [],
    model: "openai:gpt-5-mini",
    system_prompt: """
    You are a friendly greeter.
    Generate a short, warm welcome message.
    One or two sentences maximum.
    """

  def test do
    {:ok, pid} = Jido.AgentServer.start_link(agent: GreeterAi)
    GreeterAi.ask_sync(pid, "Say hello to me")
  end
end

On running,

> GreeterAi.test()
{:ok, ""}

I did not have openai credits left, so should've been an error.
I confirmed with sync call -

> Jido.AI.ask("Say Hi", model: "openai:gpt-5-mini")

And it did respond correctly with error tuple. That led to finding this bug.