PCH vs. type canonicalization

[Jason Merrill]

I'm running into a bug which only appears with PCH; it turns out to be a
problem whereby build_function_type isn't finding a version of the type
loaded from the PCH, because the hash calculation uses the tree address of
the return type.

I suppose the right fix is to stop using the address in calculating the
hash of a type.  Alternately, we could do something along the lines of the
reorder functions for sorted field structures.

Jason