MAINT: Changed the NBitBase variancy in number from co- to invariant

[BvB93]

As was noted in #18128 (comment), the currently covariant nature of np.number[~NBit_co] caused the likes of np.float32 to be a sub-type of np.float64, behavior that is very much undesirable.

This PR attempts to fix aforementioned issue by making np.number invariant w.r.t. its numerical precision.
Fortunately, this variancy change could be made quite safely and only a single test needed actual updating (08b32c3).

Examples

The behavior prior to this PR.

import numpy as np

def func(a: float64) -> None: ...

f2 = np.float16()
f4 = np.float32()
f8 = np.float64()

# These all pass without error as they are sub-types of `float64` (a.k.a. `floating[_64Bit]`)
func(f2)
func(f4)
func(f8)

[BvB93]

Pinging @seberg and @a-reich.
Continuing from #18128 (comment), it turns out that the np.number variancy could be changed without issue.

[a-reich]

@BvB93 Nice! Looks like having number invariant is not too bad after all, if that was the only substantial change. I thought e.g. arithmetic operators might need to get messier, but this version is just as clean (mixed-precision ops returned Unions before anyway so it's just a different Union EDIT: whoops I forgot that with covariance mypy simplifies the Union to just one dtype - sorry).
If this gets merged, then having ndarray covariant in dtype makes more sense. (I guess users could still do incorrect assignments on arrays of abstract dtype, but that's fine and will probably be rare).

[charris]

Needs rebase.