[1.8.x] Optimized allow_lazy() by not generating a new lazy wrapper on each invocation.

This dramatically improves performance on PyPy. The following benchmark:

python -mtimeit -s "from django.utils.functional import allow_lazy; from django.utils.translation import ugettext_lazy; f = allow_lazy(lambda s: s, str)" "f(ugettext_lazy('abc'))"

goes from 390us per loop to 165us.

Backport of 82e0cd15 from master