Nolearn

----------------

- convnet: Add scikit-learn estimator based on Jia and Donahue's
`DeCAF`.

- dbn: Change default args of `use_re_lu=True` and `nesterov=True`.

----------------

- dbn: Add parameters `learn_rate_decays` and `learn_rate_minimums`,
which allow for decreasing the learning after each epoch of
fine-tuning.

- dbn: Allow `-1` as the value of the input and output layers of the
neural network. The shapes of `X` and `y` will then be used to
determine those.

- dbn: Add support for processing sparse input data matrices.

- dbn: Improve miserable speed of `DBN.predict_proba`.

------------------

- Added a scikit-learn estimator based on George Dahl's `gdbn` in
`nolearn.dbn`.