Biggus allows indexes which aren't allowed by numpy

[djkirkham]

>>> import biggus
>>> import numpy as np
>>> a = np.arange(5)
>>> b = biggus.NumpyArrayAdapter(a)
>>> c = b[((1,),)]
>>> c.ndarray()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File ".../biggus/_init.py", line 1455, in ndarray
    array = self._apply_keys()
  File ".../biggus/_init.py", line 1525, in _apply_keys
    array = self.concrete.__getitem__(keys)
IndexError: unsupported iterator index

This is the root of a hard-to-track-down bug in Iris, so it would be useful if an error was returned when the array was first indexed.

[djkirkham]

Interestingly, for a numpy array, the following access is fine:

a[[[1]]]

but not for a biggus NumpyArrayAdapter:

>>> b[[[1]]]
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File ".../biggus/_init.py", line 516, in __getitem__
    indexed_array = self._getitem_full_keys(keys)
  File ".../biggus/_init.py", line 1444, in _getitem_full_keys
    result_key = self._cleanup_new_key(new_key, size, axis)
  File ".../biggus/_init.py", line 1358, in _cleanup_new_key
    raise IndexError(msg)
IndexError: index [1] is out of bounds for axis 0 with size 5

And the following is fine for numpy:

a[[1]]

and biggus only fails at the call to b.ndarray:

>>> b[[1]]
<NumpyArrayAdapter shape=(1) dtype=dtype('int64')>
>>> b[[1]].ndarray()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File ".../biggus/_init.py", line 1455, in ndarray
    array = self._apply_keys()
  File ".../biggus/_init.py", line 1525, in _apply_keys
    array = self.concrete.__getitem__(keys)
IndexError: unsupported iterator index

[djkirkham]

The initial issue is due to a numpy bug which is fixed in the latest version. Of the other two cases, the second (a[[1]]) is due to the same bug, while the first (a[[[1]]]) appears to be due to the fact that biggus expects arguments to __getitem__ to only be nested at most twice (e.g. 1, (1,) and ((1,),) are fine but (((1,),),) isn't), whereas numpy allows indefinite nesting