All sorts of PRs over the past two weeks.

Most importantly, initial support for new adapters, with far more smooth support for Chat LMs.

The default behavior is unchanged from before. However, dspy.configure(experimental=True) is likely to lead to much more predictable and accurate zero-shot generations with Chat LMs. Tested against several variants of GPT-3.5, GPT-4, Llama3, Mixtral, and DBRX.