Dendron

Big update this time. Here are the highlights:

- *HFLM*. We're using a new back end for calling into LLMs. HFLM is a library that makes it easy to download and use models from the Hugging Face hub.
- *New Node Types*. Thanks to increased functionality provided by HFLM, we have new node types for LLM functionality: `GenerateAction`, `LogLikelihoodAction`, and `LogLikelihoodRollingAction`. These allow you to generate text and compute probabilities of text strings and multiple choice completions. (Duncanswilson)
- *Model Centralization*. We have moved model code from individual nodes to their enclosing `BehaviorTree`, which makes it easier to reuse the same model in multiple nodes across a tree. (RichardKelley)

The above changes have necessitated API modifications, examples of which you can find in the `examples` folder. We're currently working on updating the tutorial to reflect the new API, and soon we'll be releasing a new tool for creating behavior trees more easily.

Dendron should start to work better on CPU-only systems.

- Avoid loading bitsandbytes when not quantizing models in CausalLMAction (dansbecker).
- Avoid loading bitsandbytes when not quantizing models in ImageLMAction, CompletionCondition (RichardKelley).

Getting Dendron playing nicely with Apple systems is now on the horizon.

Added support for emitting log probabilities from a `CompletionCondition` node.

Added support for ViP-LLaVa. In so doing, created a mechanism to easily add support for new vision-language models in the `ImageLMAction` node.

Updated language model nodes to properly use BitsAndBytes configuration. This should remove deprecation warnings.