feat: use previous_response_id for Codex API optimization

[Skyline-23]

Summary

Implements previous_response_id support to significantly reduce response latency when conversations grow long with Codex models.

Fixes #9045

Changes

• Add responseId field to AssistantMessage schema to persist response IDs
• Store responseId from providerMetadata.openai.responseId in processor.ts
• Add previousResponseId to StreamInput type in llm.ts
• Pass previousResponseId through providerOptions in transform.ts
• Extract last assistant's responseId in prompt.ts and compaction.ts

How it works

Before: Every request sends the entire conversation history, causing O(n) or worse latency growth.

After: Subsequent requests reference the previous response via previous_response_id, allowing the API to skip re-processing the conversation history.

Request 1 → Full history → responseId saved
Request 2 → previousResponseId + new content → Fast response

Testing

• Build passes successfully
• The feature only activates for OpenAI providers that return responseId in metadata