fix(core): batch record parallel tool outputs to prevent partial history state

[Galaxy-0]

Problem

When processing parallel tool calls, drain_in_flight was recording output items one by one as they completed. This created a race condition where the conversation history could be in an inconsistent state (containing a tool call without its corresponding output) if inspected by a background task (like auto-compact or seatbelt) during the draining process. This resulted in "tool_call_id missing response" errors (GitHub Issue #8479).

Solution

Modified drain_in_flight in core/src/codex.rs to:

1. Collect all completed tool outputs from the current batch into a Vec.
2. Record them to the conversation history in a single atomic record_conversation_items call.

Validation

• cargo check -p codex-core passed.
• cargo clippy -p codex-core passed.
• cargo test -p codex-core passed.