[PATCH] [analyzer][Review request] Fix for PR18394.

[Jordan Rose]

  Ah, right, because the destructor call happens //before// the `delete`. We should probably still do //both,// for future-proofing.

  Since there might be PreImplicitCall nodes between the previous statement and the CallEnter node, that should probably be included in the condition. We need to make sure we keep the first node before all of them. (This should basically mirror the logic in ExprEngine::replayWithoutInlining, which makes me wonder if it should be factored out.)

  Also, don't forget to include the test case in the commit.

http://llvm-reviews.chandlerc.com/D2899