Allow previous_response_id to be passed to input guardrails

[rm-openai]

Input guardrails are expected to run on the full input, so passing the previous_response_id along. For back-compat, introducing a new signature for the guardrail. Using a dataclass for the inputs so we can extend in the future easily.

[pakrym-oai]

Did you consider adding a field to RunContextWrapper instead?

[rm-openai]

@pakrym-oai yes. Decided not to, because RunContextWrapper is the same across multiple LLM calls in a single run (and hence the previous response ID changes)

[rm-openai]

After offline discussion: previous_response_id is spreading all over, and we should do this better. Workaround for end-users for now, is to add the ID to the context and use it in guardrails.