Added support for multi-process local LLM

[bshikin]

Add TestMultiprocessLocalLLM, a class that can wrap an LLM implementation and run inference (aka query method) in separate processes on the local box.

The configuration would look like:

llm:
  class: MultiprocessLocalLLM
  llm_config:
      class: <wrapped LLM class>
      config <wrapped LLM config>

    wrapper_config
      num_workers: 5,
      query_timeout_sec: 10