[C++ PATCH] Fix placement new if operator new has 2nd argument REFERENCE_TYPE (PR c++/34862)

[Jason Merrill]

It seems to me that this (and Ian's original patch) are approaching this 
problem the wrong way: we shouldn't be trying to guess whether a pointer 
is changing type based on the form of the call to new; this won't handle 
a user-defined placement new, or a user-defined override of the 
non-placement operator new which always returns the same pointer.  Can't 
we attach the CHANGE_DYNAMIC_TYPE_EXPR to the value returned from 
operator new, rather than mess with the arguments?

Jason