Add a better mutually_broadcastable_shapes strategy

[asmeurer]

This is based on the work from
Quansight-Labs/ndindex#126. The strategy that comes
with hypothesis has default parameters that makes it not generate interesting
examples by default. In particular, it will never generate a shape with more
than 2 dimensions and side length more than 3.

This new strategy generates shapes that are based on a base shape drawn from
shapes(), so more complex shapes are possible. This is preferable to doing so
by modifying the parameters to mutually_broadcastable_shapes() because it's
difficult to find parameters that give interesting examples but don't require
too much filtering for total size.

This commit also changes it so that mutually_broadcastable_shapes generates
shapes with minimum side length 0 by default.

For example, before this commit:

>>> from array_api_tests.hypothesis_helpers import mutually_broadcastable_shapes
>>> for i in range(10):
...     print(mutually_broadcastable_shapes(2).example())
((3,), (3, 1))
((), (3,))
((), (2,))
((1,), ())
((3,), ())
((), (1,))
((1,), (1, 1))
((2,), ())
((), (1,))
((3,), ())

After this commit:

>>> for i in range(10):
...     print(mutually_broadcastable_shapes(2).example())
((), (1, 1))
((0, 1), (1, 1, 1, 1))
((1, 3, 1, 1), (3, 1, 4))
((), ())
((1, 5, 1), (1, 5, 1))
((1, 1), (2, 2))
((1,), (2, 0, 1))
((), (1, 1, 1, 2))
((), ())
((), (1, 4, 1))

The generation with this new strategy is somewhat slower than the old one
(about 2x slower), but the generation speed is not slow enough that it should
matter too much.

The shrinking can also be imperfect with this strategy. For example, I've seen
it shrink to ((0, 1), (0, 2)) for the currently failing in-place tests where
it could shrink to just ((1,), (2,)). I also saw the pre-existing strategy
have the same problem, however. One should just be aware that test failures
involving shapes from this strategy might not actually be shrunk minimally.

[honno]

Generating base shapes is a nice idea. Shame about minimisation, would rather solve this now as opposed to later, but it is something I can mull over.

Carrying over kwargs to base_shape is blocking (see review comment), otherwise this LGTM.

[honno]

We can merge this, but it's worth just checking if changing the defaulting does what you want—please try out #75. This would be preferable for minimisation behaviour, as well as be less of a maintenance burden.

Just printing the results of a typical run, it seems quite comparable. I don't have much practical experience with the strategy like you do though.

[asmeurer]

The simpler change #75 should work for now. If we ever decide to modify shapes from the default strategy, this may end up being a better approach. Ultimately, I'd prefer if hypothesis allowed filtering both shapes and mutually_broadcastable_shapes by total size in a more direct way.