Pickling of multivarlinreg objects

[saroele]

This pull requests solves pickling issues with patsy in the multivarlinreg.
Solution: with __getstate__() and __setstate__() we can modify the object to be pickled and restore it.
Implementation details as described in the docstring of __getstate__() :

    Remove attributes that cannot be pickled and store as dict.

    Each fit has a model.formula which is a patsy ModelDesc and this cannot be pickled.
    We use our knowledge of this ModelDesc (as we build it up manually in the do_analysis() method)
    and decompose it into a dictionary.  This dictionary is stored in the list 'formulas',
    one dict per fit.

    Finally we have to remove each fit entirely (not just the formula), it is built-up again
    from self.formulas in the __setstate__ method.

ping @JrtPec who needs this functionality. Maybe you can check and merge this PR?

[coveralls]

Coverage increased (+1.05%) to 92.027% when pulling 7fd46a5 on issue35 into 7a91d29 on master.

[coveralls]

Coverage increased (+1.05%) to 92.027% when pulling 7fd46a5 on issue35 into 7a91d29 on master.