Fix meta for reduction scalars

[fjetter]

So, it turns out that dask/dask#10671 is not necessarily a problem. The problem why the reduction test originally failed is because the Scalar object had a python bool as it's meta instead of a numpy bool. Dask's assert logic has a special path for numpy dtypes that implements a more forgiving equality check I ran into here.

I suspect this is something we'll run into many times when comparing with the dask/dask test suite. This feels like something make_meta should deal with already properly. I'll poke around dask/dask if we can fix it at the source.

[fjetter]

hmm this fix is not working well it seems. I think dask/dask#10677 will be better