decrement N in FreqDist.pop and FreqDist.popitem, remove arg from FreqDist.popitem; resolves issue 726

fix cosine_distance for KMeansClusterer 

see issue 709 http://code.google.com/p/nltk/issues/detail?id=709
in short, cosine_distance is implemented as a similarity metric,
while KMeansClusterer expects a distance metric.

Optional argument for KMeansClusterer and new implementation for centroid calculation, to avoid empty clusters

Accessing self._en_wordlist raises AttributeError on first run

The `if not self._en_wordlist` clause does not protect from the AttributeError
that is raised when the user attempts to access the non-existent attribute. 

reworking import statements to avoid wildcard imports; making nltk.sem.util depend on fewer external modules