In the error case, we will clear the GVariant and then try to emit a
response from a NULL pointer as result. Always create an empty vardict
when we something failed instead.

This also clears the error because we continue on in the error case and
might re-use it. We do in fact re-use it in one case which is a
programmer error.

By g_variant_ref_sink'ing floating GVariants as early as possible. This
fixes a few cases where the variant would be leaked, where the floating
ref would be owned by a g_autoptr and makes it easier to see who owns
the reference.

Cases where the floating ref is passed into a function which immediately
sinks the floating ref are left untouched.