feat: add SWE-bench fullset support

[xingyaoww]

What is the problem that this fixes or functionality that this introduces? Does it fix any open issues?

---

Give a summary of what the PR does, explaining any non-trivial design decisions

• Support setting the SWE-Bench dataset and split to run evaluation using command line argument.
• Update the documentation to allow people to directly run_infer.sh without running a pull instance docker first -> run_infer.sh should be able to automatically pull them.
• Add script with documentation for pushing new SWE-Bench images to remote registry.

---

Other references

[yufansong]

QQ: does the evaluation need so many images? 🙀

[xingyaoww]

Each instance per image, so yes 😢 - that's why we need a good infra to run this at scale

[yufansong]

Fine, that is crazy.

[jatinganhotra]

Thanks @xingyaoww . This is exactly what I was looking for...

    if args.set == 'full-test':
        dataset = load_dataset('princeton-nlp/SWE-bench', split='test')
    elif args.set == 'lite-test':
        dataset = load_dataset('princeton-nlp/SWE-bench_Lite', split='test')

It would be awesome if you could add 'princeton-nlp/SWE-bench', split='dev' and 'princeton-nlp/SWE-bench_Lite', split='dev') as well :)

[xingyaoww]

Hey @jatinganhotra, I tried to look into this by adding dev set -- but was blocked by princeton-nlp/SWE-bench#199. Will try to add support for dev again once that issue is resolved.

[jatinganhotra]

Being addressed by #3478

[neubig]

Looks good, thanks!