BUG: `timedelta64` does not support boolean (floor-)division

[BvB93]

Despite booleans being castable to timedelta, the np.true_divide and np.floor_divide ufuncs
currently do not support divisions between np.timedelta64 and np.bool_.
Note that timedelta / integer operations do work as expected.

This is inconsistent w.r.t. the rest of the ufuncs, which support either no timedelta casting or the full range of casts:

• np.add and np.multiply support the full range of timedelta casts,
  i.e. both timedelta + timedelta, timedelta + integer and timedelta + bool operations are accepted.
• np.remainder (ENH: timedelta64 mod and floordiv #16195) and np.nat do not support timedelta casting at all.

Reproducing code example:

Division:

In [1]: import numpy as np

In [2]: b = np.array([True])
   ...: u = np.array([1], dtype=np.uint)
   ...: i = np.array([1])
   ...: m = np.array([1], dtype="timedelta64[s]")

In [3]: m / i
Out[3]: array([1], dtype='timedelta64[s]')

In [4]: m / u
Out[4]: array([1], dtype='timedelta64[s]')

In [5]: m / b
---------------------------------------------------------------------------
UFuncTypeError: ufunc 'true_divide' cannot use operands with types dtype('<m8[s]') and dtype('bool')

Floor division:

In [6]: m // i
Out[6]: array([1], dtype='timedelta64[s]')

In [7]: m // u
Out[7]: array([1], dtype='timedelta64[s]')

In [8]: m // b
---------------------------------------------------------------------------
UFuncTypeError: ufunc 'floor_divide' cannot use operands with types dtype('<m8[s]') and dtype('bool')

NumPy/Python version information:

In [9]: import sys, numpy; print(numpy.__version__, sys.version)
1.21.0.dev0+807.g121b6097a 3.8.5 (default, Sep  4 2020, 02:22:02) 
[Clang 10.0.0 ]

[seberg]

@BvB93 do we really want that? Note that they can cast "safely", but they do not promote in general (or at least I started deprecating that 😈. In my world view, that would mean, I would rather make that + ufunc here use "promotion" (not casting), which would effectively deprecate the above.
But, I admit that timedelta64 + integer is probably intentionally allowed, even if dubious.

Saying all of that, timedelta64 / boolean is actually like timedelta64 / integer which is valid (and not the same as timdelta64 / timedelta64 due to the unit!). Since we allow upcasts of boolean to integers everywhere, division specifically should probably be just allowed.

[BvB93]

@BvB93 do we really want that?

Well, what I'm mostly interested in is consistency, and np.floor_divide and np.true_divide are currently (as far as timedeltas + bool ops are concerned) inconsistent w.r.t. to all other ufuncs, an inconsistency that, in my opinion, appears to be unjustified.

I'm a bit less concerned with how this could be resolved, be it either via allowing or dissallowing mixed timdelta/boolean operations in general. Though as you've noted in the division example, I don't see any harm in allowing it.