This now will allow to add a max_inter option to agents while also making sure to force the agent to give it's best final answer before running out of it's max_inter.
* Refactoring task cache to be a tool
The previous implementation of the task caching system was early exiting
the agent executor due to the fact it was returning an AgentFinish object.
This now refactors it to use a cache specific tool that is dynamically
added and forced into the agent in case of a task execution that was
already executed with the same input.
* Adding tool caching a loop execution prevention.
This adds some guardrails, to both prevent the same tool to be used
consecutively and also caching tool's results across the entire crew
so it cuts down execution time and eventual LLM calls.
This plays a huge role for smaller opensource models that usually fall
into those behaviors patterns.
It also includes some smaller improvements around the tool prompt and
agent tools, all with the same intention of guiding models into
better conform with agent instructions.
