Cast was not needed since the return value was already being cast. Query
function changed to use heapq.nlargest() for simplicity

Problems fixed:
Replaced cmp: 1 instance in line 323 of search.py
Convert set() to {}: 3 instances in line 29 of MDP.py, line 181 and 215
of search.py
Convert .sort() to sorted(): 1 instance in line 150 of text.py
Cast map to list: 1 instance in line 47 of grid.py

- Fix relative imports in base of is a package or not
- Fix not working notebooks
- Add notebooks to packages sources

Used tool 2to3 to finish the porting to Python 3 plus some fixes by
hand.

These functions, which I had to implement in Python 2.2, are now
available in standard Python.

Move CountingProbDist to learning.py and use it in NaiveBayesLearner. Make NaiveBayesLearner use target-value frequencies too.

 * Use weighted_sampler() for the sampling.
 * By not using a defaultdict, keep n_obs always up to date.
 * Don't inherit from ProbDist, at least for now, since we don't
   use anything from the superclass, and if we did it would break
   (e.g. __setitem__).
 * Remove __len__ since nobody uses it.
 * Use heapq.nlargest.

Also separate random doctests from others, and tweak docs.