feat: forward chat_id as user field in OpenAI-compatible chat completions

The helper _inject_chat_id_as_user calls metadata.get('chat_id') without checking that metadata is a dict. If a client sends a non-dict value for metadata (e.g. an empty list [], a string, a number), metadata.get(...) raises AttributeError, resulting in a 500 error. The condition only uses truthiness (if metadata), which passes for non-empty lists and other truthy non-dict objects. This introduces a new crash path compared to the previous code that just popped and ignored metadata.

_inject_chat_id_as_user reads ENABLE_FORWARD_CHAT_ID from the module's import-time snapshot rather than from env.py on every call. This means _inject_chat_id_as_user will never observe a runtime change to the environment variable — it behaves as a static constant. While this matches the default-deferred resolution pattern used elsewhere in this codebase (e.g., ENABLE_FORWARD_USER_INFO_HEADERS), it breaks unit-test isolation: the test suite monkeypatches the router module instead of the source-of-truth env module, creating a divergence between test behavior and production. The pattern is fragile; future maintainers may re-import and lose the monkeypatch. Consider reading the env variable lazily inside the helper or during request processing so that runtime changes are reflected and tests can set the environment directly without patching module state.