… of skip_axes

This changes the way the strategy is generated to construct skip_axes and a
result shape first, then generate shapes that match those skip axes and
broadcast to the result shape. It doesn't attempt to directly use
mutually_broadcastable_shapes. This is unfortunate, and it does require a
little bit of filtering, but it is the only way that I could figure out how to
generate the full range of possible examples. skip_axes_st is now always
generated in conjunction with this strategy, as 1) generating the skip axes
first and then the shapes doesn't seem to work, and 2) the two strategies are
always used together anyway.

This doesn't yet update all the tests that use this (nor are the functions
themselves updated), but test_mutually_broadcastable_shapes_with_skip_axes()
passes.

…th skip_axes

This is impossible to do in general if we allow both negative and nonnegative
skip_axes, and it's also too complex to check the result shape skip axes to
make sure they are consistent.

Unfortunately, this is a backwards compatibility break, but hopefully no one
is actually using this function yet.

This also changes the definition of the internal associated_axis() function.

The testing strategy is changed a little bit. We just test against the
broadcasted arrays with the skip axes removed, since the number of skip axes
is not necessarily uniform across the shapes anymore.

Only generate slices for completely skipped shapes once.

It is (hopefully) actually correct now with the new definition for skip_axes.

…index into iter_indices-multi-skip_axes