codex-mcp: serialize connector runtime refreshes

[mzeng-openai]

What

• Add one async explicit-refresh lock per active connector runtime context.
• Add McpConnectionManager::hard_refresh_codex_apps_runtime, which performs exactly one uncached Apps tools/list under that shared lock and returns the exact committed Arc<ConnectorRuntimeSnapshot>.
• Start the host-owned Apps MCP client in initialize-only mode for the new strict path so startup cannot issue a second competing list request.
• Keep legacy hard_refresh_codex_apps_tools / app/list behavior intact while routing its refresh through the same lock.
• Retain the previously committed snapshot when fetch or publication fails.
• Record the uncached list outcome/trigger and expose the committed snapshot to core for later consumers.

Why

This is stack PR 2 of 4, based on #31471. It creates the single refresh seam needed by request-stable runtime consumption and app/installed without changing app-server APIs yet.

Invariants

• Concurrent explicit reloads serialize per account/workspace context.
• A strict reload performs one network tools/list, then atomically publishes and returns that same committed state.
• Generic tool filtering, exposure, approval, elicitation, and execution remain owned by McpConnectionManager.
• The legacy app/list compatibility path remains available and retains its existing response behavior.

Checks

• just test -p codex-mcp (114 passed)
• just fmt

The broader codex-core suite was also compiled and exercised while developing this slice; its connector/MCP tests passed, with only the checkout's known sandbox/CLI-fixture failures in the full crate run.