c++: Fix up convert_nontype_argument [PR124070]

c++: Fix up convert_nontype_argument [PR124070]

Some of the following testcases are incorrectly accepted.
The problem is in convert_nontype_argument.  It just does
maybe_constant_value earlier.  Later on for integral/enum/floating
point args it uses cxx_constant_value for complain & tf_error if
not val_dep_p and expr is not a simple constant, similarly for
class type it calls
      /* Replace the argument with a reference to the corresponding template
         parameter object.  */
      if (!val_dep_p)
        expr = create_template_parm_object (expr, complain);
      if (expr == error_mark_node)
        return NULL_TREE;
etc.  But for NULLPTR_TYPE_P and REFLECTION_TYPE_P cases (I bet
Marek has based the latter on the former) it only verifies the type
and not the actual expression.
I think we need to do that, because if maybe_constant_value fails, it
just returns the passed in argument and if nothing checks it, we accept
whatever was there (and sometimes ICE later).

The following patch implements that.

2026-02-19  Jakub Jelinek  <jakub@redhat.com>

	PR c++/124070
	* pt.cc (convert_nontype_argument): For NULLPTR_TYPE_P case
	if expr is not integer_zerop and is not val_dep_p try harder,
	for complain & tf_error use cxx_constant_value and return NULL_TREE
	for error_mark_node, without tf_error return NULL_TREE.  Similarly
	for REFLECTION_TYPE_P case if expr is not REFLECT_EXPR and is not
	val_dep_p try harder like for NULLPTR_TYPE_P.

	* g++.dg/cpp26/pr124070.C: New test.
	* g++.dg/reflect/parameters_of4.C: Expect different diagnostics.
	* g++.dg/reflect/define_aggregate7.C: New test.
	* g++.dg/reflect/define_aggregate8.C: New test.