fine-grained: strip is_abstract

[msullivan]

Fixes a crash bug

[ilevkivskyi]

Good catch!

This gives me two ideas:

• From time to time we can run fine grained tests while randomizing order of targets reprocessing and running every test few dozen times. I have been bitten few times by this, and it was quite painful. And I bet if we will do this now, then we could discover few another similar bugs.
• We can order target modules not by name (which is quite arbitrary) but by topological sort (plus some heuristics within cycles similar to what we currently have to order modules within SCCs in normal incremental). I already proposed this from the point of view of performance, but now it looks like stability/predictability may be a more important argument.

[msullivan]

I actually just hacked something up to do a fairly basic randomization (just randomizing the order of files passed in), and after accounting for differing error message order and one test that has two possible files that can block it, I haven't been able to get a failure out of it, which I find pleasantly surprising.

Possibly randomizing the order of target reprocessing might cause more grief