Fix handling of thread features for scalars in Anderson2021

[aekul]

This fixes the issue in #7699. The cause is that the autoscheduler assumes that every Func is surrounded by a block loop and a thread loop, which means it assumes there is always a non-null ThreadInfo object when the features are computed. But in this case, the output of the generator is a single number so the autoscheduler is currently not creating a block and thread loop for it, so the ThreadInfo object ends up being null.

The main thing this PR does is add some special cases to compute_features for handling scalars. It was also necessary to change some function signatures to accept a ThreadInfo* instead of a ThreadInfo& so that it can test whether the passed object is null or not.

[steven-johnson]

LGTM, but I presume that #7703 is still an issue -- after this lands, do we want to consider #7706 as a fix for it, or do you have an alternate suggestion?

[steven-johnson]

(Oh, and thanks SO much for taking the time to look at this!)

[steven-johnson]

This looks like it may have injected #7729 based on a quick bisection -- I thought that anderson2021_test_apps_autoscheduler had run on the testbots, but maybe I misread due to other false-positive failures :-/

[aekul]

No problem. Thanks for looking into the test failure.

And yes, I think we should consider #7706 as a fix for #7703