Currently deserialization of a non-packed binary repeated field is quadratic in
the number of elements, because each time we parse a new element we copy over
all elements we have parsed so far. This CL fixes the performance problem by
having the generated deserialization code just call addX() instead of using
getX() and setX().