previous_response_id stateful chain support

[HXYerror]

Part of #1. Depends on #3.

Goal

OpenAI's Responses API supports two ways to carry context across turns:

1. Stateless: the client echoes prior reasoning items (with encrypted_content) into input (covered by Model-to-endpoint routing (chat vs responses) #5)
2. Stateful: the client passes previous_response_id: \"resp_…\" and only sends the new user turn; the server reconstitutes the prior context

We should support both. (1) is the path Codex CLI prefers for ZDR; (2) is what some lighter Responses clients use.

Current state

previous_response_id is unmentioned anywhere in the repo (Grep 0 hits, see #1 audit).

Tasks

• Add previous_response_id?: string to the request type (Add upstream Responses API service client #3)
• Pass it through to upstream verbatim — copilot-api does not need to maintain its own state store; Copilot upstream handles it
• Add id: string (the response id) to the response type if not already there from Add upstream Responses API service client #3
• Document in the README that both stateful and stateless multi-turn modes work
• (Optional, future) If Copilot upstream's state retention turns out to differ from OpenAI's (e.g., shorter TTL, different store semantics), surface that in docs

Acceptance criteria

• Two-turn test: turn 1 returns id: \"resp_X\". Turn 2 sends only the new user message + previous_response_id: \"resp_X\". Reply correctly references context from turn 1.

File pointers

• OpenAI docs: https://platform.openai.com/docs/guides/conversation-state

[HXYerror]

Duplicate of the corresponding #12 — closing to avoid confusion.