add gpt-4o support for num_tokens

[nicklamiller]

Without support in num_tokens for the gpt-4omodel, running OpenAIChatCompletionRequest with model=gpt-4o results in a Failure:

requests = [OpenAIChatCompletionRequest("what's the answer to life?", model="gpt-4o")]
responses = list(chat_completion_batch(requests, label="prompt"))

responses[0].response
# Failure(_value=(OpenAIChatCompletionRequest(messages=[{'role': 'user', 'content': "what's the answer to life?"}], model='gpt-4o', temperature=0.8, top_p=1.0, n=1, stop=None, max_tokens=None, presence_penalty=0.0, frequency_penalty=0.0, logit_bias=None, metadata=None), NotImplementedError('num_tokens_from_messages() is not implemented for model gpt-4o. See https://github.com/openai/openai-python/blob/main/chatml.md for information on how messages are converted to tokens.')))

Section 6 of the OpenAI cookbook for counting tokens doesn't lay out token counting rules for gpt-4o like it does for the other models (checked main for newest version of the notebook). I tested a few messages and got the same results for both gpt-4 and gpt-4o so it would appear they count tokens the same, though it would be good to confirm this.

Testing tokens are same between gpt-4 and gpt-4o

from oai_utils.openai_utils import num_tokens


all_example_messages = [
    	[{"role":"user", "content": "hi"}],
    	[{"role":"user", "content": "a longer message"}],
    	[{"role":"user", "content": "a much longer message indeed"}],
]
for example_messages in all_example_messages:
    for model in ["gpt-4", "gpt-4o"]:
        print(model)
        print(f"{num_tokens(example_messages, model)} prompt tokens counted by num_tokens().")
        response = openai.ChatCompletion.create(
            model=model,
            messages=example_messages,
            temperature=0,
            max_tokens=1,  # we're only counting input tokens here, so let's not waste tokens on the output
        )
        print(f'{response["usage"]["prompt_tokens"]} prompt tokens counted by the OpenAI API.')
        print()

[ravwojdyla]

lgtm! In general maybe we should take the snapshot on the internal OpenAI tooling and dump it in here again, aka "snapshot" style open source. Out of scope of this PR.