feat: use token limits in core memory instead of character limits

[cpacker]

2000 tokens is a good starting point? (~8k chars)

Pending design decision:

• Do we make the Block.tokenizer_model nullable / Optional[str], or str with a default imported from constants.py?
  • For example, if the field is nullable, in the Pydantic validator that throws an error if the token count is higher than the limit, we can just use constants.DEFAULT_TOKENIZER_MODEL if self.tokenizer_model is None
  • Or (if we set defaults), self.tokenizer_model will be constants.DEFAULT_TOKENIZER_MODEL
  • Seems to me like the key difference is if we make the field nullable / have the default lazily inferred, then we can change the default for all existing blocks with new package versions?
• Or do we put tokenizer_model inside of metadata_?

TODOs for later PR:

• Initialize the tokenizer_model in the Memory class to be drawn from agent.llm_config.model
• Make sure that when the agent.llm_config.model is changed, the memory.tokenizer_model is changed too