_newscalar creates a scalar value of type typ with specified dtype.
Inside it works via creating dtype that will have its .type = typ and
that's how NumPy knows to instantiate scalars of such dtype via
specified typ. But at the end _newscalar was checking only that .dtype
of creatged scalar == original dtype, while we can insist that
scalar.dtype _is_ the created dtype which itself should be == to
originally passed dtype.

No change in behaviour, only make the logic more robust and explicit.