c++: Fix ICE with #embed/RAW_DATA_CST after list conversion [PR118671]

The following testcases ICE with RAW_DATA_CSTs (so the first one since
introduction of #embed C++ optimizations and the latter since optimization
of large sequences of comma separated literals).
I've missed the fact that implicit_conversion can embed the exact expression
passed to it into stuff pointed out by conversion * (e.g. for user
conversions in sub->cand->args).
So, it isn't enough in convert_like_internal to pass the right INTEGER_CST
for each element of the RAW_DATA_CST because the whole RAW_DATA_CST might be
in sub->cand->args etc.
Either I'd need to chase for wherever the RAW_DATA_CST is found and update
those for each element processed, or, as implemented in the following patch,
build_list_conv detects the easy optimizable case where
convert_like_internal can be kept as the whole RAW_DATA_CST with changed
type and possibly narrowing diagnostics, and otherwise instead of having
a single subconversion it has RAW_DATA_CST separate subconversions.
Instead of trying to reallocate the subconvs array when we detect that case,
the patch instead uses an artificial ck_list inside of the u.list array
to hold the individual subconversions.
Seems the only places where u.list is used are build_list_conv and
convert_like_internal.

2025-02-04  Jakub Jelinek  <jakub@redhat.com>

	PR c++/118671
	* call.cc (build_list_conv): For RAW_DATA_CST, call
	implicit_conversion with INTEGER_CST representing first byte instead
	of the whole RAW_DATA_CST.  If it is an optimizable trivial
	conversion, just save that to subconvs, otherwise allocate an
	artificial ck_list for all the RAW_DATA_CST bytes and create
	subsubconv for each of them.
	(convert_like_internal): For ck_list with RAW_DATA_CST, instead of
	doing all the checks for optimizable conversion just check kind and
	assert everything else, otherwise use subsubconversions instead of
	the subconversion for each element.

	* g++.dg/cpp/embed-25.C: New test.
	* g++.dg/cpp0x/pr118671.C: New test.