c++: Fix deduction from the type of an NTTP

c++: Fix deduction from the type of an NTTP

In the testcase nontype-auto17.C below, the calls to f and g are invalid
because neither deduction nor defaulting of the template parameter T
yields a valid specialization.  Deducing T doesn't work because T is
used only in a non-deduced context, and defaulting T doesn't work
because its default argument makes the type of M invalid.

But with -std=c++17 or later, we incorrectly accept both calls.
Starting with C++17 (specifically P0127R2), during deduction we're
allowed to try to deduce T from the argument '42' that's been
tentatively deduced for M.  The problem is that when unify walks into
the type of M (a TYPENAME_TYPE), it immediately gives up without
performing any new unifications (so the type of M is still unknown) --
and then we go on to unify M with '42' anyway.  Later in
type_unification_real, we complete the template argument vector using
T's default template argument, and end up forming the bogus
specializations f<void, 42> and g<S, 42>.

This patch fixes this issue by checking whether the type of an NTTP is
still dependent after walking into its type during unification.  If it
is, it means we couldn't deduce all the template parameters used in its
type, and so we shouldn't yet unify the NTTP.

(The new testcase ttp33.C demonstrates the need for the TEMPLATE_PARM_LEVEL
check; without it, we would ICE on this testcase from the call to tsubst.)

gcc/cp/ChangeLog:

	* pt.c (unify) <case TEMPLATE_PARM_INDEX>: After walking into
	the type of the NTTP, substitute into the type again.  If the
	type is still dependent, don't unify the NTTP.

gcc/testsuite/ChangeLog:

	* g++.dg/template/partial5.C: Adjust directives to expect the
	same errors across all dialects.
	* g++.dg/cpp1z/nontype-auto17.C: New test.
	* g++.dg/cpp1z/nontype-auto18.C: New test.
	* g++.dg/template/ttp33.C: New test.