Add scikit-learn style random_state

[pavlin-policar]

Fixes #29.

• Adds numpy as a dependency for the nice RandomState facilities. This shouldn't be too problematic since most everybody has numpy installed.
• check_random_state is copied from scikit learn for consistent behaviour. I thought adding scikit-learn as a dependency was overkill for this simple function.

randomize now behaves like this:

• True will produce completely random results, same as the current behaviour
• False will set the random seed to 0, thus producing the same results each time. This is not the same as the current behaviour, which does not permute the node order at all.
• I've changed the default value to be None. Setting it to anything other than None will result in a deprecation warning. In future versions, this parameter should be removed so there aren't two parameters that control randomness.

random_state behaves in exactly the same way as in scikit-learn.

By default, we now produce random result, which is consistent with scikit-learn. Before the default was to return deterministic results.

If you don't like any of this, I can change it, but I don't think any of this is problematic. It's better to be consistent with popular libraries so the API is familiar, and I've added deprecation warnings so users can gradually switch to the new API.

[taynaud]

Sorry it took me a while.

I will try to make a few more checks this week end and do a release.

Thanks!