Fix _assert_shared

[djkirkham]

This fixes the issue that now the _assert_shared function can't detect when a pandas object shares its data with a numpy array.

I'm not really happy with this solution. There's nothing in the pandas documentation that says values returns the original data object (indeed, in 0.16.2 it doesn't). But I can't find another way.

As an aside, a note in the doctring of _assert_shared as_series states

.. note::

   Pandas will sometimes make a copy of the array,
   for example when creating from an int32 array.
   Iris will detect this and raise an exception if copy=False.

This doesn't appear to be the case any more. At least, values always seems to return the original array, and the buffer objects of a pandas.Series and the NumPy array from which it was created always seem to point to the same location in memory, if cope=False is specified.

[marqh]

please add a comment around this change, explaining why the change is needed and referencing pandas versions as appropriate

I would like to see just enough about the pandas implementation change to be able to understand the change if I came back to it

[marqh]

this looks fine
thank you @djkirkham