DOC: Rewrite docstrings of `ogrid` and `mgrid`

* BUG: properly handle tuple keys in NpZFile.__getitem__

* TST: test tuple rendering specifically.

---------

Co-authored-by: Ross Barnowski <rossbar@berkeley.edu>

Removed references to `nd_grid` instead of adding a document (gh-23612).

This roughly changes things so that we ensure a float64 working
values up-front.  There is a tiny chance of precision changes if the
input was not float64 or error changes on bad input.

I don't think this should matter in practice, precision changes
(as far as I can tell) should happen rather the other way around.

Since float64 has 53bits mantissa, I think the arange should give
the correct result reliably for any sensible inputs.

There is an argument to be made that the windowing functions could
return float32 for float32 input, but I somewhat think this is OK
and users can be expected to just cast manually after the fact.

The result type is tested, but this ensures the tests pass also
when enabling weak promotion.

This PR adds additional tests for quantiles:

1. Identification equation $E[V(q, Y)] = 0$
2. Adding a constant $c > 0$: $q(c + Y) = c + q(Y)$
3. Multiplying by a constant $c > 0$: $q(c \cdot Y) = c \cdot q(y)$
4. Multiplying by $-1$: $q_{\alpha}(-Y) = -q_{1-\alpha}(Y)$

(Note by seberg as reviewer: These tests are fairly complex, but may be useful for future development.  But if they seem too complicated, they are probably not really vital and could be shortened or removed.)

DOC: quantile parameter q is a probability

NpzFile inherits from collections.abc.Mapping,
which provides __contains__().
However, it calls __getitem__(),
which can be slow because it performs file decompression on success.