feat: Respect backoff policy for streaming

[timonv]

Went into the rabbit hole of streaming backoffs. Turns out eventsource has something already, of which I'm not 100% certain how well it works, but since streaming is a bit different, we want to generally retry all the time. This PR sets the event source retry policy to the (similar) values of backoff. There's one slight difference in behaviour, in that on transport errors, we only retry server errors or 429s, like with regular requests.

We can't inspect the response body at this point, so we can't differentiate between rate limits and quota errors (thanks openai).

Also, during streaming, token errors are a rough 400. I think maybe. Sometimes.

Practically, I'm trying to stream a lot of agents in parallel, hoping with this I can mitigate it somewhat.

[timonv]

Actually, #413 would be a nice extension to this

[timonv]

@64bit My approach here still feels a bit like a hack. EventSource I feel is the real challenge here, perhaps replacing that or otherwise provides a more robust solution.

For reference, I solved it in swiftide by hooking into the stream itself. This works really well, we use it in production at scale.

A custom streaming backoff policy: https://github.com/bosun-ai/swiftide/blob/master/swiftide-core/src/stream_backoff.rs#L31
Applying the strategy to clients via a decorator: https://github.com/bosun-ai/swiftide/blob/master/swiftide-core/src/indexing_decorators.rs#L131

What are your thoughts on this? #428 seems to be gaining some traction. I haven't been able yet to get gpt5 to stream properly.