Support for cross-platform serialization of MLP.

[alexjc]

The weights are now pickled instead of the MLP entirely. Additional objects such as input_space are required to _create_mlp() once the de-pickling happens. Some tests in place, should be added to Classifier and Regressor tests too.

Updates #7.

[alexjc]

Things to do: determinism test for all activation types comparing pre- and post-serialization — Maxout in particular.

[coveralls]

Coverage decreased (-1.44%) to 98.56% when pulling 4343da5 on serialization into a57ac0c on master.

[coveralls]

Coverage decreased (-1.44%) to 98.56% when pulling 4343da5 on serialization into a57ac0c on master.

[coveralls]

Coverage remained the same at 100.0% when pulling d0c09f0 on serialization into a57ac0c on master.

[coveralls]

Coverage remained the same at 100.0% when pulling d0c09f0 on serialization into a57ac0c on master.

[coveralls]

Coverage remained the same at 100.0% when pulling 9c33047 on serialization into a57ac0c on master.

[alexjc]

Ready for review @ssamot when you have time :-)

[alexjc]

Closes #7.

[coveralls]

Coverage remained the same at 100.0% when pulling 8c32d48 on serialization into a57ac0c on master.

[alexjc]

Addressed @ssamot's comments by adding a test to ensure that numpy.ndarrays are serialized to disk and not theano data-structures, ensuring platform compatibility.