Add a span processor collect in-progress OpenTelemetry traces during execution and pass them to the LLM mocker. For example, in a calculator evaluation run that calls the get_random_operator tool twice, the LLM mocker gets the following run history added to its context.

First LLM mock:

CURRENT AGENT RUN SO FAR:
(empty)

Second LLM mock:

CURRENT AGENT RUN SO FAR:
get_random_operator
  Input: {}
  Output: {"result": "+"}