Add CustomAdapter for evaluation tools

[andrewtran117]

Adds CustomAdapter for onboarding user defined custom evaluation tools. Provides defaults for required schema fields and supports JSON/JSONL/directory inputs.

[borgr]

Overall it looks good, I mainly wonder about higher level stuff here (naming and usecases).
So currently this works specifically from jsons, maybe the naming should convey it?
It is not custom in the sense that you expect someone to change the code to read their files. Someone still needs to change the outputs to the jsons, when will people use it?
I am thinking for example now on the EvEvEv people, will they want to use this code somehow? (They parse leaderboards into our format). We need to add to README and documentation to lead people to use it when they have a relevant usecase.