feat: added basic heartbeat override heuristics

[cpacker]

Closes #601

---

Please describe the purpose of this pull request.

Local/open LLMs can often forget to request a heartbeat when running functions. E.g.:

> Enter your message: Using your archival memory, what do you know about Banh Mi?
💭 The user has asked me to use my archival knowledge to find information about 'Banh Mi'. This is a task I can handle easily. I will now search for all the memories containing that term.
⚡🧠 [function] updating memory with archival_memory_search
        query: banh mi, page: 0
> Enter your message:

This PR adds a basic heuristics to manually insert the heartbeat request:

• If the user sends a message (type == user_message), and the agent immediately runs a function (non-send-message function), 99% chance the user is expecting a reply

How to test

• Basic test parser (pytest -s -vvv tests/test_function_parser.py)
• Test with old agents
• Test with new agents
• Test with OpenAI (regression)
• Test with local LLM

Have you tested this PR?

Yes, see below.

---

Side note: had to patch a new bug in LM Studio where the prefix is no longer adhered to (probably to do with their default prompt formatting changing?), which requires unpacking a nested function due to our json patchers:

[2023-12-15 12:11:29.967] [INFO] [LM STUDIO SERVER] Generated prediction: {
  "id": "chatcmpl-",
  "object": "chat.completion",
  "created": 1702671084,
  "model": "/Users/loaner/lmstudio_models/TheBloke/OpenHermes-2.5-Mistral-7B-16k-GGUF/openhermes-2.5-mistral-7b-16k.Q8_0.gguf",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "{\n  \"function\": \"send_message\",\n  \"params\": {\n    \"inner_thoughts\": \"Activating persona Sam. Persona is designed to be curious, empathetic, and self-aware.\",\n    \"message\": \"Hello Chad! I'm Sam, your new companion. Nice to meet you!\"\n  }\n}"
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 2908,
    "completion_tokens": 71,
    "total_tokens": 2979
  }
}

[cpacker]

Trying to patch out weird behavior on function returns:

what do you know about me?
actually my name is brad (short for bradley)
trigger name update

Example mixtral (hosted endpoint, temp = 0.8)

Before changes, model doesn't understand that the control got returned due to the heartbeat and goes off-topic / hallucinates:

Testing on mixtral (hosted endpoint, temp = 0.8)

New heartbeat str versions seem to be working

Testing on local dolphin 2.2.1

Also seems to be working

Testing on local openhermes 2.5

Bad func call but return is good