Correct branch folding for HCR guard versioning

This logic has been inadvertently folding the versioning test instead of
the original guard since ce45970, which
is effectively the same as not versioning the guards. This can be
expected to have resulted in slower code than intended. Worse, if there
are HCR guards in the loop, then once the fixed-point analysis has
determined that it is possible to version them, they must be versioned
for correctness. This is because other transformations may rely on the
cold side being eliminated from the hot loop. In particular, by failing
to version HCR guards, we could privatize mutable fields even though
they could be modified by a cold call while the loop is still running.

compiler/optimizer/LoopVersioner.cpp

@@ -4585,7 +4585,7 @@ void TR_LoopVersioner::versionNaturalLoop(TR_RegionStructure *whileLoop, List<TR
          comparisonTrees.add(guard);
 
          bool reverseBranch = false, origLoop = true;
-         FoldConditional fold(this, NULL, guard, reverseBranch, origLoop);
+         FoldConditional fold(this, NULL, tt->getNode(), reverseBranch, origLoop);
          fold.improveLoop();
          }
       }