The former has been formally deprecated for a long time and in CUDA 10.0
it is finally properly marked as deprecated, so the compiler can print
warnings.

General reductions on host

See merge request mmg/tnl-dev!7

According to benchmarks, there is practically no difference in
performance. Only explicit unrolling is helpful, but that has been
implemented for the general algorithm in Reduction::reduce as well.

Now that error handling is done via exceptions, the return value is free
for more sensible things.