ACCKernelsTrans should exclude character assignments

[arporter]

Daley has found that if we include character assignments within a KERNELS region, the nvfortran compiler (24.5) errors with:

Character local variable unsupported on device

Unfortunately, a simple example does not replicate this problem which indicates it's probably a compiler bug and depends on the precise context.

NEMO source files that exercise this problem: obs_group_def.f90, sbccpl.f90

[arporter]

I think the sensible thing for us to do is to exclude character assignments. In #2602 I did initially alter ACCKernelsTrans to exclude all character literals but that had a big, negative performance impact. Therefore, I refined it to check for character vars of assumed size. I can try just checking for literals on the RHS of assignments...

[arporter]

We could have an option to permit them so that a developer can always enable them if the compiler is happy.

[arporter]

A similar but related problem...we get an ICE in lib_mpp (the nml reading code therein) for:

          !$acc kernels
          cdnambuff(itot:itot + inl - 1) = chline(1:inl)
          !$acc end kernels

where:

   CHARACTER(LEN = :), ALLOCATABLE, INTENT(INOUT) :: cdnambuff

and the error is:

NVFORTRAN-S-0000-Internal compiler error. size_of: attempt to size deferred size character

[arporter]

Sergi has pointed out that we don't have this problem in the OMP-related transformation script because there, we first use the ArrayAssignment2LoopsTrans transformation and the validate of that class checks for string assignments:

PSyclone/src/psyclone/psyir/transformations/arrayassignment2loops_trans.py

Lines 288 to 304 in 2307ac7

	for child in node.walk((Literal, Reference)):
	try:
	forbidden = ScalarType.Intrinsic.CHARACTER
	if (child.is_character(unknown_as=False) or
	(child.symbol.datatype.intrinsic == forbidden)):
	message = (f"{self.name} does not expand ranges "
	f"on character arrays by default (use the"
	f"'allow_string' option to expand them)")
	if verbose:
	node.append_preceding_comment(message)
	# pylint: disable=cell-var-from-loop
	raise TransformationError(LazyString(
	lambda: f"{message}, but found:"
	f"\n{node.debug_string()}"))
	except (NotImplementedError, AttributeError):
	# We cannot always get the datatype, we ignore this for now
	pass

It's not a lot of code so I think the simplest solution is to copy that bit into the validate of ACCKernelsTrans. I think any other transformation will be OK because they too will have to first use ArrayAssignment2LoopsTrans.