openai sse: cap the pending-line buffer in SseState #152

SseState (crates/tinyinference-llm/src/providers/openai/sse.rs) appends every received chunk to buf and never limits the unterminated line it holds. An endpoint that sends a long data: line without a newline can grow the buffer until the stream ends or the request times out.

This affects every streaming caller of the OpenAI-compatible adapter and predates #151, which only routes local-openai invokes through the same parser. Raised in review there (CodeRabbit).

Suggested fix: a fixed maximum for a pending line (for example 1 MiB), checked after each chunk is drained; when it is exceeded, end the stream with a non-retryable ProviderFailed, as the out-of-range tool-call index guard does. Add a test with a synthetic oversized line.