feat: respect registry rate limits

We should add a rate limit checks to building images to avoid overwhelming docker.

A generic 429 Too Many Requests error, or a specific "Too Many Requests (HAP429)" error is generally referring to the abuse limit, described here: https://docs.docker.com/docker-hub/usage/#abuse-rate-limit
This limit applies to all requests to Hub properties including web pages, APIs, and image pulls. The limit is applied per-IP, and while the limit changes over time depending on load and other factors, it's in the order of thousands of requests per minute. The abuse limit applies to all users equally regardless of account level.

To avoid hitting this limit accidentally we should try to keep the api calls and image pulls to under 100 a minute.
Assuming we use 100 workers, this shouldn't delay evaluations. Even if we use 500 workers - unlikely due to LLM rate limits - it would only delay it by 4 minutes. Keeping it under 100 prevents accidentally reaching the 1000 limit when running multiple evaluations in parallel

This feature affects SWE-Bench, commit0, and SWT-bench and all feature benchmarks.

GAIA implements its own rate limit logic until (https://github.com/OpenHands/benchmarks/issues/171) is implemented.

Start by locating the image-building, registry API, and image-pull entry points used by SWE-Bench, commit0, SWT-bench, and feature benchmarks. Review how GAIA currently limits requests and confirm the intended behavior for parallel evaluations; done means requests stay under the stated limit without delaying evaluations unnecessarily.