ci: retry transient Docker pulls during integration setup

Two integration jobs failed during setup when Docker Hub timed out while pulling images:

- https://github.com/apache/iceberg-go/actions/runs/30167050157/job/89701755026
- https://github.com/apache/iceberg-go/actions/runs/30167050070/job/89701754817

Both failed before tests ran. Their Spark 4 counterparts passed.

Integration setup currently runs docker compose up -d --wait once, so a temporary registry timeout fails the whole job.

A few possible fixes:

1. Retry docker compose pull in CI with backoff, then start with --pull never. This seems safest and does not change local behavior.
2. Retry the full integration setup command. This is simpler, but may recreate a partially started environment.
3. Limit concurrent image pulls and add backoff to reduce registry pressure.
4. Use a registry mirror or preloaded image cache.

Would you prefer the CI-only pull retry, or should the retry behavior live in the Makefile so it can also be used locally?

Happy to send a PR with the preferred approach.

Start by inspecting the CI integration setup that runs `docker compose up -d --wait` and review the two linked failed job logs. Choose and document the retry scope with the maintainers; done means transient Docker Hub pull failures no longer fail setup while local behavior remains unchanged.

Written by the indexing model from the issue text.