Fixes https://github.com/cython/cython/issues/3663

This ensures that rvalues here are saved as temps, while keeping the
existing behaviour for `for x in deref(vec)`, where the pointer for vec
is copied, meaning it doesn't crash if vec is reassigned.

The bit of this change liable to have the biggest effect is that I've
changed the result type of dereference(x) and x[0] (where x is a c++
type) to a reference rather than value type. I think this is OK because
it matches what C++ does. If that isn't a sensible change then I can
probably inspect the loop sequence more closely to try to detect this.