openai SSE: decide whether an undecodable stream chunk should fail the call #153

Raised in review on tinyhumansai/tinyagents#610. Today an SSE chunk that doesn't deserialize as a ChatCompletionChunk is skipped with a WARN (tinyinference#150). The alternative is to end the stream with a provider failure. Trade-off: failing surfaces silent data loss loudly, but it turns any provider's harmless odd frame (extension frames, new fields) into a failed call. Needs a decision on whether to fail on chunks that look like content, or stay lenient with a metric.