Demangler: further speed improvements.

Avoid using std::string and std::vector in the demangler. Instead use vectors/strings with storage allocated by the NodeFactory’s bump pointer allocator.
This brings another 35% speedup. Especially in the case the Demangle::Context is not reused for subsequent demanglings.

include/swift/Basic/Demangle.h

@@ -207,6 +207,9 @@ public:
 
   // Only to be used by the demangler parsers.
   void addChild(NodePointer Child, NodeFactory &Factory);
+
+  // Reverses the order of children.
+  void reverseChildren(size_t StartingAt = 0);
 };
 
 /// Returns true if the mangledName starts with the swift mangling prefix.