Add suppression for input and output data

[cartermp]

Adds two config parameters:

• suppress_response_data, which when set to True, will NOT log response data (like chat responses) to a span
• suppress_input_content, which when set to True, will NOT log input data (like a corpus of text passed to a model with a large context window) to a span

This lets people avoid reaching limits on spans. Imagine an app that passes in a large amount of data per request, or a long conversation in a chatbot where the entire conversation history is passed in as input each time -- this would likely exceed the total size a span can have.

[estib]

Looks good -- although i am interested in talking through one question:

Will people be able to predict when content / responses are likely to be too long at the time they are initiating the openai OTEL auto instrumentor? Or will it be something a bit more dynamic, like "sometimes i get a few really long questions / responses, but sometimes not" depending on the users?

What do you think about adding a config to define dynamic rules of when a response_data or input_content should be included? Like maybe we could let users define some kind of length limit for how long a response_data should be before it gets suppressed? This could be something we do in addition to a boolean on/off switch like what this PR adds, but it could also be something we do instead? (could even allow users to set the length limit to 0 to suppress them entirely)

I dunno, what do you think?

[cartermp]

Mmmm, this is a good question:

Will people be able to predict when content / responses are likely to be too long at the time they are initiating the openai OTEL auto instrumentor? Or will it be something a bit more dynamic, like "sometimes i get a few really long questions / responses, but sometimes not" depending on the users?

You can certainly predict it for single requests, or at least a range of I/O size. But yeah, for genuine chat apps, there's no predicting things, nor is there predicting things for agents where you continually build larger and larger context.

I think a setting that's more dynamic in setting, limiting total inputs and/or outputs by a threshold does make sense.