This addresses #4666 and #4945.

Currently we trust retry-after when a 429 comes around and we usually handle it with retries in the SDK first. This for instance means that when you run out of opencode go quota, you get stuck there retrying forever because the timeout is measured in days.

This makes three major changes:

1. it disables sdk internal retries by default. Given the fact that we're already doing agent level retries that is the most sensible
2. it enforces a 60 second upper bound for retry-after which is what other systems also do
3. it sniffs the error messaage and tries to detect out of usage 429s and does not retry them

For opencode go it now looks like this instantly instead of getting stuck:

Error: 429 Monthly usage limit reached. Resets in 16 days. To continue using this model now, enable usage from your available balance: <snip>