REL: Prepare for the NumPy 1.22.0rc2 release.

- Update 1.22.0-notes.rst.
- Update 1.22.0-changelog.rst.

REV: Revert adding a default ufunc promoter

We can't trust them! Just don't... if they can be cached it does
not matter anyway...

Adding a default promoter should not be necessary with the
reduction-hack (use output DType to guess loop DType if incompatible
with the loop).

However, using it had two effects:
1. SciPy has a ufunc that has a loop it does not want to exist.
   This loop is homogeneous (unlike the correct one) and gives
   a deprecation warning.  The default promoter would assume the
   homogeneous loop is OK (maybe not ideal) if it exists.
   Since it is a "bad" loop that gives a deprecation warning, it
   is not really true though.
2. Datetime promotion is currently utterly buggy, leading to:
       timedelta.sum(dtype="f8")
   ignoring the `f8` requests.  But we actually end up relying on
   that behaviour in our `np.mean` implementation...

REL: Prepare for NumPy 1.22.0rc2 release.

- Update 1.22.0-changelog.rst.
- Update 1.22.0-notes.rst.

PERF: Fix performance bug in dispatching cache

BUG: Fix handling of the dtype parameter to numpy.array_api.prod()

In promotion cases, the result of the dispatching was not cached
(at least not as well as it could be).  This lead to large slowdowns
in cases where promotion is necessary, one example:

    a = np.array([3]); b = np.array([3.])
    %timeit a + b

The fix here is fairly small, and almost a cleanup, since I put the
whole `cache=True` logic in the wrong place really.
(I currently do not ever cache promoters, to be fair, this could
be done but should be unnecessary and may require a bit of thought.
Another thing that could be optimized is caching if there is no
matching loop to speed up error paths.)