Fix GVariant handling in regress_test_closure_variant

Fix GVariant handling in regress_test_closure_variant

Commit 7c4fbbd6 introduced some low-level g_memdup()ing of the passed GVariant
in regress_test_closure_variant(). This causes a race condition with GI clients
like pygobject's test suite, which sometimes fail with

  (runtests.py:15653): GLib-CRITICAL **: g_variant_ref_sink: assertion `value->ref_count > 0' failed

Drop the memduping again, and instead make the argument non-const (which is
true, as the method temporarily increases the refcount).