Use raw input instead of prepared input to input guardrail

[aaron-ang]

Summary

This change ensures that input guardrails receive the original user input rather than the input combined with session history. Input guardrails should validate the user's current message, not the entire conversation context, as they are designed to check if the new input is appropriate rather than validating the full conversation history.

In addition, I created some helper functions that abstract duplicated chunks of code across run and _start_streaming methods.

Checks

• I've added new tests (if relevant)
• I've added/updated the relevant documentation
• I've run make lint and make format (skipped file changes since they are unrelated to the PR)
• I've made sure tests pass

[rm-openai]

Passing the input history was intentional. Some input guardrails only operate on the newest message, and they can just filter the inputs. But others operate in the context of the overall conversation, and this would be a regression for them.

(LMK if there's a use case that's not covered, happy to discuss how we can enable it)