Shake up `TestUtils.run_ad` interface again

[penelopeysm]

1. We need to get rid of the VarInfo argument. Given that you can just pass a transform strategy instead, it serves very little use now...

apart from

allowing people to stick values inside it with unflatten, and then those values become the values that are used for gradient preparation. We actually use that in ADTests to demonstrate that ReverseDiffCompiled fails sometimes. There are two ways around this:

1. Declare that this usage is just way too advanced, and unnecessary. We can hardcode the ADTests table to show that it's wrong.
2. Put another keyword argument on LogDensityFunction, that allows users to control the parameters used for gradient prep, and then put a keyword argument on run_ad that forwards to the LDF.

Finally, we need to:

• get rid of the varinfo field in the ADResult (arguably it should just straight up store the LDF)

That is easy to do, but it's breaking, so it will have to go in the next minor version.

[anurag-mds]

I'd like to work on this. My plan is to remove the varinfo parameter from run_ad and replace the three-branch initialization block with the transform_strategy path only, then update ADResult to store the LogDensityFunction instead of varinfo. Does this approach align with what you had in mind?

[penelopeysm]

Yes, largely. Please make a PR against the breaking branch though -- this is an API change so will have to be released as a minor version. (https://turinglang.org/docs/developers/contributing/#pull-requests-versions-and-releases)

[anurag-mds]

@penelopeysm I have created the PR review it and do let me know if anything needs to be changed

[penelopeysm]

Closed via #1311