first performance improvements for SLIM

[remiadon]

Describe your proposed solution

Two simple things to do:

• right now SLIM make sure not to drop itemsets with zero usages when switching a codetable for a new, more concise one. This results in a final codetable with a lot of empty usage lists. Decide where to drop those zeros, and where to let them will lead to better run time and user experience
• Like mentionned in the litterature, when considering a new candidate, reduce the database to itemsets containing all items in this itemset. This will make computation of covers faster

[remiadon]

mentioned in #8

Binary representation may help us:

• computing covers faster than with a naive iterative algorithm
• improving inference time, i.e runtimes for .predict_* methods, which is linked to computing covers faster

[remiadon]

Added to this : present a use case on 20news group