[COMMITTED PATCH] Avoid C++ tests when the C++ cannot be linked.

[Roland McGrath]

> I don't like this. The semantics I would rather have are that all such
> tests fail, not fail to build, but fail. Then the results of `make check`
> represent reality with all the C++ tests failing because you don't have
> a libstdc++. This way it's clear that you haven't passed all of the testing
> for glibc.

I don't like the conflation of unbuildable with failed.  What we need is
the possibility of UNSUPPORTED as an outcome for our tests.  I'll send a
change in that direction.