fix(llm): omit stateless response item ids

[rekram1-node]

Summary

• omit provider item IDs when replaying complete OpenAI reasoning items with store: false
• retain item IDs internally for summary-block grouping and provider metadata
• preserve stateful item_reference continuations when storage is enabled

Rationale

The Open Responses input schema makes ReasoningItemParam.id optional, while ItemReferenceParam.id remains required. OpenAI Codex follows that distinction by skipping response item IDs during normal request serialization and only attaching them for its stateful Azure path.

Stateless continuations already carry reasoning.encrypted_content, so they should replay the complete item without asserting the identity of a provider-stored item. This complements #34013, which selects encrypted replay for catalog OpenAI models.

Tests

• bun run test test/provider/openai-responses.test.ts test/tool-runtime.test.ts: 78 passed
• bun run test: 281 passed, 30 skipped
• bun typecheck
• bunx prettier --check packages/llm/src/protocols/openai-responses.ts packages/llm/test/provider/openai-responses.test.ts packages/llm/test/tool-runtime.test.ts
• bunx oxlint packages/llm/src/protocols/openai-responses.ts packages/llm/test/provider/openai-responses.test.ts packages/llm/test/tool-runtime.test.ts: zero errors, one pre-existing warning in unchanged test code
• git diff --check