Disallowing timedelta64+float promotion (to timedelta64) in all cases
(previously it was assymetric and "half allowed") meant that isclose,
allclose, np.ma.allclose, and assert_arrays_almost_equal (which uses
isclose), would stop work for timedelta64. Hardcoding that timedelta64
is passed on unmodified retains the old behaviour.

It may make sense to deprecate or change this behaviour in the future,
but for the 1.20 release, the behaviour should be as much unmodified
as possible.

Closes gh-18286

... or when the input isn't/cannot be a set.  I left a few usages, e.g.
in random sampling, where "set" is reasonable as informal description of
an array as the order doesn't matter; however, for e.g. np.gradient the
order of the returned list is clearly important, so "set" is wrong.

Also some other minor doc edits noticed during the grepping: using
`shape` instead of `form` in `cov` is consistent with most other places;
the wording in `Polynomial.trim` now matches other methods on the same
class.

DEP: deprecate np.testing.dec

Since the decorators in np.testing.dec are intended for nose test framework
support, deprecate them by updating the docstring and adding a warning when
they are used on functions

This update a coupe of references (single backticks) that actually are not to
verbatim/code (double backticks); and a couple of verbatim to reference
when they do actually exists and can be resolved in context.

I probably missed other; and stayed simple but spoted a few other
inconsistencies that I did not fix:

  - some ``...`` could actually be :math:`...` but not always clear if
  it would be better.
  - some intervals are [``...``], other are ``[...]``

I guess they could be discussed individually; it was mostly the failing
references that bothered me.