Accept types in internal CSV reading function

[folmos-at-orange]

Description

When solving issues #74 and #128, we realized that when reading internal files, the table types were computed automatically by pandas.

For predict, this wasn't a problem since we cast to the proper type in the sklearn post-processing of a khiops execution. However, now that we want to accept more types there are some edge-cases (eg. a string column of True and False values) that are not well handled. The problem is compounded with the fact that we accept the "meta-type" pd.Categorical whose internal representation may be any of the basic types.

Questions/Ideas

• Make the read_khiops_internal_file accept a parameter column_dtypes
  • Specialize for each leaf class the reading code for predict, predict_proba and transform. Currently it is at the KhiopsEstimator level.