Add whole-program inference mode using stub files instead of jaif files

[kelloggm]

This PR adds a command-line options to support using .astub files as the output of whole-program inference instead of .jaif files, and permits -Ainfer to take an argument:

• -AmergeStubsWithSource causes stub types to be used, even when source files are available; the stronger type from the two is used.

For now, there is scant documentation on -Ainfer=stubs, because there is no practical way for a user to make use of it (for now). This is because the manual only describes how to run whole-program inference via the infer-and-annotate.sh script, which is not practical for large programs. Providing a version of infer-and-annotate.sh that uses stubs (or something more scalable) is explicitly future work for this PR.

The stub mode of WPI is tested with the same tests that the .jaif version of whole-program inference uses. Those tests also implicitly test -AmergeStubsWithSource, because it's not possible to overwrite (a lack of) annotations in source files otherwise.

Merge together with typetools/annotation-tools#279

[smillst]

The commit history is a bit funky because I merged an earlier version of #2846 and built off of that. Those commits claim they've been merged, but they're still showing up in the diff here. I've tried getting rid of them, and they're required for my tests to pass locally. @smillst, were those merged? If not, can I merged them here? Or do they need separate tests?

I think I fixed this. Can you double check the diffs?

[kelloggm]

Yes, that works now. Thanks, I appreciate it.

It looks like some of the all-systems tests are failing; I'll look into that and then re-assign you when tests are passing for review.

[smillst]

I haven't looked at all the code yet. Could you please add a comment for every method, even private ones?

[smillst]

Use {@code } instead of the html tag.

[smillst]

I would change -AoutputStubs to -Ainfer=stubs. Then this would be checker.getOption("infer").equals("stubs"). That makes documenting this option easier.

[smillst]

Replace with a TypeScanner:

        new SimpleAnnotatedTypeScanner<Boolean, Void>() {
            @Override
            protected Boolean defaultAction(AnnotatedTypeMirror type, Void aVoid) {
                return type.getAnnotations().isEmpty();
            }

            @Override
            protected Boolean reduce(Boolean r1, Boolean r2) {
                r1 = r1 == null ? false : r1;
                r2 = r2 == null ? false : r2;
                return r1 || r2;
            }
        };

[kelloggm]

I've added javadoc to all the methods and fields, even the private ones, and made the changes @smillst suggested (with one exception: I was able to avoid the need to scan for the presence of annotations in general, so I could remove the ugly string hack entirely without replacing it with a SimpleAnnotatedTypeScanner). @smillst can you take another look?

[smillst]

@kelloggm Please update the pull request description to match the implementation.