Support for boolean indexing

[esc24]

Biggus is used in Iris. Before the switch to use Biggus one could used boolean indexing on cubes (see https://github.com/SciTools/iris/pull/643). Biggus does not currently support this ability e.g.

>>> import biggus
>>> import numpy as np
>>> a = biggus.NumpyArrayAdapter(np.arange(12).reshape(3,4))
>>> a
<NumpyArrayAdapter shape=(3, 4) dtype=dtype('int64')>
>>> a[np.array([True, False, True])]
<NumpyArrayAdapter shape=(3, 4) dtype=dtype('int64')>

This should be

<NumpyArrayAdapter shape=(2, 4) dtype=dtype('int64')>

This PR adds this functionality.

[esc24]

Looks like this needs a little extra work as

a = biggus.NumpyArrayAdapter(np.arange(12).reshape(3,4))[:, :]
a[np.array([True, False, True])]

gives an unexpected answer.

[pelson]

Just checking you know about:

>>> import numpy as np
>>> a = np.arange(12)
>>> a[np.zeros(12, dtype=np.bool)]
array([], dtype=int64)
>>> a[np.zeros(10, dtype=np.bool)]
array([], dtype=int64)
>>> a[np.ones(10, dtype=np.bool)]
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

[esc24]

Thanks @pelson. I'd missed that. I've relaxed the restriction so the key is effectively zero padded if it is short. Overall, that was more complicated than I'd expected, but I've learnt something.

[pelson]

:+1: - I would merge. My only hesitation would be to test the actual index values, not just the shapes.

I'll give this a couple of days before merging.

[rhattersley]

the key is effectively zero padded if it is short

This behaviour might have changed in NumPy 1.9.

[rhattersley]

This PR adds [boolean indexing].

Super! :smile: