Speed up importing

[akx]

As discussed in #1069, importing tflearn can be very slow due to it importing all of tensorflow.contrib.

This PR significantly improves import speed:

(master) $ time python -c 'import tflearn'
        5.45 real         4.73 user         0.54 sys
(fast-import) $ time python -c 'import tflearn'
        1.37 real         1.17 user         0.19 sys

This is achieved by lazy-loading any tensorflow.contrib components that would otherwise be always loaded, with some caveats:

• The first use of tflearn.initializations.xavier and .variance_scaling will be slow
• tflearn.data_utils.VocabularyProcessor no longer inherits from tensorflow.contrib.learn.python.learn.preprocessing.text.VocabularyProcessor, which might break some isinstance() calls in user code.
• The tensorflow/contrib/framework/python/ops/arg_scope.py file is vendored in tflearn.

Regarding VocabularyProcessor, though – it looks like the class did nothing more than wrap calls to its superclass, so maybe it should be removed altogether with a note to have people use the superclass directly instead (though it too is deprecated since TensorFlow 1.7.0).

Closes #1069

[aymericdamien]

Great! Thanks for helping with that. Importing contrib was indeed slowing down quite a lot.