performance: parallelize diffgraph application

[bbrehm]

Can you @mpollmeier test that on the same thing you use to evaluate ArrayList vs ArrayBuffer?
Can you test this on a big cpg creation?

(the issue with cpg creation is temp memory during diffgraph application. If we parallelize, then more stuff is alive at the same time. I also had to remove some early clearing of memory, because that would require more synchronization to figure out when an object is truly dead)

[mpollmeier]

thank you, will do 👍🏻

[bbrehm]

Ok, it wasn't so bad to restore the early cleanup logic (so that the GC can reclaim structures that are not needed anymore, even while the diffgraph application is still running).

Main things to consider are:

1. The atomic refcount rigamarole
2. The Callable closure is kept alive by the java.util.concurrent.ForkJoinTask.AdaptedCallable after it finished, until we do submissions.clear(). So we need to be careful about what we capture in the closure!

[mpollmeier]

import timing results: #286 (comment)

[mpollmeier]

Won't merge for now, because the performance gains currently don't justify the added complexity. But this might come in handy in future, so let's keep the PR open.

Context:
#286