fix(usage): accept MiniMax-style chat.completion final SSE frame

[nonoqing]

MiniMax (and possibly other OpenAI-compatible providers) close a streaming response with a non-streaming chat.completion frame instead of a true chat.completion.chunk. That final frame is the only one carrying the authoritative usage block — earlier chunks send only zero placeholders. Before this fix BitFun dropped the frame at two layers:

1. The weak validator in stream_handler/openai.rs required the object string to be exactly "chat.completion.chunk" and labeled MiniMax's final frame as skip:non_standard_event, so it never reached deserialization.
2. Even if the validator had let it through, the choice in that frame uses message instead of delta, and OpenAISSEData required Choice.delta to be present.

The net effect: BitFun recorded 0 input, 0 output, 0 cached for every MiniMax call. Fix both layers:

• Validator now accepts both chat.completion.chunk and chat.completion object strings.
• Choice.delta is now #[serde(default)] with Delta deriving Default, so frames lacking delta parse cleanly. We don't need the frame's content (earlier chunks streamed it); we only need top-level usage and finish_reason to propagate.

Each fix has a regression test reproducing MiniMax's observed on-the-wire shape captured from a live MiniMax-M2.7-highspeed response.