Allow the use of OpenAI's rough token count

[Sergio-e]

Context

Noticed that the tiktoken_ruby dependency isn't thread-safe. For us, this resulted in the Sidekiq process crashing consistently. There is also an open issue regarding this problem

Proposal

Allow the use of OpenAI's rough token count instead of using tiktoken ruby. The token count selection will be made when initializing the LLM:

    Langchain::LLM::OpenAI.new(
      api_key: openai_access_token,
      llm_options: {},
      default_options: { token_counter: :openai }
    )

tiktoken_ruby will be left as the default option.

[andreibondarev]

@Sergio-e Do we still need this? I got rid of the automatic max_tokens validation via tiktoken. I think that feature was more annoying than helpful.

[Sergio-e]

@Sergio-e Do we still need this? I got rid of the automatic max_tokens validation via tiktoken. I think that feature was more annoying than helpful.

It seems not. The issue is now fixed on the latest Tiktoken ruby release