Represent similarities as Floats instead of Doubles in LexRank

[karlhigley]

The precision of a Double isn't necessary for the simple comparisons that are used, and this saves a few bytes per graph edge (which adds up quickly).