Add support for Modal

[azliu0]

What does this implement/fix? Explain your changes.

Adds support for running the evaluation harness on Modal Labs.

Adding a --modal flag to the normal run command will run the instances remotely rather than locally. The current images resolves ~300/300 tasks on Lite and ~495/500 tasks on Verified.

Benefits of running this way:

• users don't have to set up their own infrastructure, everything runs in the cloud ☁️
• everything runs in parallel, so building and running larger numbers of instances is effortless.
• with full image caching, Lite runs in ~14 minutes (excluding sympy__sympy-11870, only ~7 minutes). Verified runs in ~7 minutes
• Modal's image caching is built-in and follows essentially the same layered abstraction as the images already in the repo

[wenting-zhao]

I am curious if Modal has the plan to fix the remaining 10/300?

[pawalt]

@wenting-zhao as of the most commit, we achieve 300/300 on lite and 495/500 on verified.

[carlosejimenez]

Thanks @azliu0 and @pawalt!
Just tested it out and it's great!

[john-b-yang]

Support added back in d1ede7d