Change hh.mutually_broadcastable_shapes() defaults to generate more interesting shapes

[honno]

Changes the defaulting of max_dims and max_side so that more interesting shapes get generated. Also defaults min_side to 0.

[asmeurer]

I guess this doesn't end up being that different for us because the hypothesis shapes strategy is also based on limiting the side length, whereas in ndindex I just use tuples(integer()) and filter across the total size.

[honno]

I guess this doesn't end up being that different for us because the hypothesis shapes strategy is also based on limiting the side length, whereas in ndindex I just use tuples(integer()) and filter across the total size.

Ah so you're saying #73 is limited as-is? Does this PR fair well? Maybe can bump up the defaults a bit more too.

My impression is that the defaults from Hypothesis might be too low, but I also think relying on st.integers().filter(<below max size>) is overkill (i.e. seemingly won't be finding interesting bugs for the big performance cost when making/manipulating arrays).