This patch is a follow up from r240560 and is a step further into trying
to mitigate the compile time performance issues in the path DeadStoreElimination -> MepDepAnalysis ->
CaptureTracking. I've basically moved the "cached numbered BB" out of the CaptureTracker.
By providing the CaptureTracker with the "cached numbered BB" instead of
computing it every time, MemDepAnalysis can use the same "cached numbered BB"
throughout its calls to AA->callCapturesBefore, avoiding to recompute it for
every scanned instruction. In the same testcase used in r240560, compile time goes
down from 2min to 30s.