Most Menpo Python 2 pickles cannot be loaded into Python 3

[jabooth]

This is because our pathlib.Path objects contain strings which need to marshalled to unicode on loading into Python 3.

The default for Python 3 loading a Python 2 pickle is to only decode valid ASCII characters to strings. If you come across an invalid character, you get this:

/.../lib/python3.5/site-packages/menpo/io/input/base.py in 
_import(filepath, extensions_map, landmark_resolver, landmark_ext_map, 
        landmark_attach_func, asset, importer_kwargs)
    802     if importer_kwargs is None:
    803         importer_kwargs = {}
--> 804     built_objects = importer_callable(path, asset=asset, **importer_kwargs)
    805 
    806     # landmarks are iterable so check for list precisely

/.../lib/python3.5/site-packages/menpo/io/input/pickle.py in
pickle_importer(filepath, asset, **kwargs)
     25     """
     26     with open(str(filepath), 'rb') as f:
---> 27         x = pickle.load(f)
     28     return x
     29 

UnicodeDecodeError: 'ascii' codec can't decode byte 0x81 in position 2: 
ordinal not in range(128)

A solution to this is to change the import encoding to latin1 in Python 3:

import pickle

with open('a.pkl', 'rb') as f:
    d = pickle.load(f, encoding='latin1')

However this is a bold change to make so I'm making this issue to think through any possible consequences.

If we decided to do this, note that we would need a Python version check to add it, as pickle.load() in Python 2.7 does not take any kwargs.

We should also bear in mind that we may head towards Python 3 only support (#721). This issue is both an example of the kind of burden we could rid ourselves of by moving sooner, and potentially something that isn't worth solving if a commitment to Python 3 is on the horizon (i.e. - rather than muddy the waters our official stance can be Python 2/3 pickles are incompatible and we can just remove this confusion down the line by some point dropping Python 2).

[jabooth]

A worrying reason why not to set the default to 'latin1':

numpy/numpy#4879

pv (numpy member)
Using encoding='latin1' probably works in many cases, but e.g. Python 2's datetime objects are not pickleable on Py3 with it.

[patricksnape]

Crikey.

[patricksnape]

Work around added by #723