NpyIter_Deallocate returns 0 for failure and 1 for success,
so the check `NpyIter_Deallocate(iter) < 0` is not correct.

(See relevant comments in gh-20069.)

MAINT: Refactor reductions to use NEP 43 style dispatching/promotion

In particular, this covers a casting error that currently cannot
be hit for normal ufuncs, because they already check casting during
the legacy dtype resolution (which is called first).

More fixups are coming, the biggest change is that the error message
is now improved when a reduction makes no sense such as
`np.subtract.reduce(np.array([1, 2, 3], dtype="M8[s]"))`
where input and output cannot have the same descriptor.

(Some more fixups still to go)

At least the reduceat path seems to have been untested before.
The slight change in code layout (added assert) is just to make
the code slightly easier to read.  Since otherwise it looks like
there is an additional `else` branch when `out` is given.

This also implements new promoters for the logical functions.
The reasoning is that any logical function can always be handled
by first casting to boolean without changing the meaning.
This is true (unless the output is object), and already effectively
used for `np.all`/`np.any`.