(autofix): Investigate/implement exponential backoff for streaming vertex ai calls

[ram-senth]

Looks like we're getting overloaded errors from anthropic on vertex, but we should be using a built-in exponential backoff retry strategy from the SDK. Investigate whether this is actually being used on the streaming calls and identify whether we need to implement one ourselves for the streaming calls

https://sentry.sentry.io/issues/6152252202/?notification_uuid=17a60c1a-a237-4557-8ea6-076b1e32ecd5&project=6178942

We need to watch out for error events and retry:
https://docs.anthropic.com/en/api/messages-streaming#error-events