Why are these changes needed?

It’s often useful to combine multiple advanced capabilities into a single agent, as OpenAI has done with their recently released GPTs and Assistants, for which function calling, code interpreter, and RAG capabilities can work together in a composable and chainable way. But AutoGen makes it hard to chain advanced capabilities because ConversableAgent obtains each agent reply from just one registered reply function. Given the mutual exclusivity among registered reply functions, the capabilities of (for example) RetrieveAssistantAgent and TeachableAgent cannot be easily chained (combined in producing a single response) today; any given reply can leverage one or the other but not both.

This PR enhances ConversableAgent to support chainable capabilities by means of a lightweight hook-registration mechanism allowing multiple capabilities to contribute to each agent reply. Leveraging this new mechanism, teachability can now be added to any agent that inherits from ConversableAgent by instantiating a Teachability object (which replaces the original TeachableAgent class) and calling its add_to_agent(agent) method. For illustration, teachability can now be added to OpenAI Assistants through GPTAssistantAgent (unmodified by this PR), successfully combining teachability in a chain with whatever capabilities the particular OpenAI Assistants have (tools including functions, code interpreter, and retrieval). Teachable OpenAI Assistants are demonstrated in a new notebook.

The original teachability tests, notebook, documentation, and blog post are all updated.

Related issue number

Closes #534, #535, #782

Checks

• I've included any doc changes needed for https://microsoft.github.io/autogen/. See https://microsoft.github.io/autogen/docs/Contribute#documentation to build and test documentation locally.
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed.