Streams one request to an SSE model API and normalizes the events (SuperPi.AI.StreamNormalizer).
Shared by the Anthropic and OpenAI providers.
Returns what SuperPi.AI.Provider.generate/2 returns, plus {:retry_after, ms} for a rate limit
or an overloaded server, which the conversation machine turns into a backoff.
Req's own retry is off. The machine owns retries (backoff, cap, durable state), and a second layer of retries would repeat a request the first one already counted.