Add test which verifies that we can un/pickle 'pd.DataFrame' which were
converted to 'np.recarray' via the '.to_records()' method.

This test is necessary because it turns out that the combination of
pandas >= 0.20.x + python2 doesn't support this functionality by
default if the data frames column names are unicodes.

Please see nexedi/erp5!1738 (comment 179298) for
additional context.

/reviewed-on nexedi/erp5!1738
/reviewed-by @jerome @klaus