Compute similarities between sparse rating vectors

[NicolasHug]

As discussed in #135, we can compute any similarity measure in two ways:

• only considering the common (existing) ratings. This is what we're doing ATM in surprise, see the doc for details.
• considering all possible ratings, replacing unknown ratings by a given value (typically, zero).

The goal here is to implement this last option for all the available similarity metrics.

This has been partially addressed in #141.

This would require a new field comon_ratings_only in the sim_options parameter, which would default to True. Also, possibly a fill field representing the value used in replacement of missing ratings.