Adds Improved Error Handling, Documentation, and Tests For The Refine Module

[cezarc1]

Of note: Adds configurable fail_count parameter to Refine to control how many errors are tolerated before throwing. By default we will throw once if all of the underlying module's calls fails.

Adds in-line documentation to both Refine and BestOfN and in the cheatsheet.

[okhat]

This is amazing, thanks so much @cezarc1 ! Timing is interesting because we're considering how to act on this and #7922 together. We will merge both PRs, but the question is what's the final design that emerges from that.

[cezarc1]

@okhat . I see! I missed #7922. As I think it was mentioned in #7922 I still see the benefits of Refine and BestofN still existing as my understanding is that the retry mechanism in that PR requires a new adapter, while Refine and BestOfN do not require a new adapter in order to work.

My $0.02 is that, from a design perspective, both Refine and BestofN (or other mechanisms) should work without any special new Adapters.

An alternative is for feedback, as a concept, should be implemented into the Adapter system, similar to how demos currently are.