Documentation about tasks <> request types mapping

[anjor]

This table in the readme outlining which request type is supported for each API is super useful.

However, as a user I usually want to run a set of benchmarks (tasks) against a set of APIs. Not knowing how the tasks map to request types means I don't know until I actually try to run the benchmark whether or not the run is going to succeed.

For example, running hellaswag with openai chat completions fails because loglikelihood and loglikelihood_rolling are not implemented due to lack of support for logits.

[haileyschoelkopf]

Will complete as part of #1160 ! This is very useful feedback.