improve Covington optimization

[chrislit]

Covington is by far the slowest algorithm in the library. There are some aspects that will be hard to optimize away, but it's worth looking through the code to see if the sim/dist functions themselves could be optimized at least.

[chrislit]

I don't see any clear method of optimizing this, so closing. The one area that might be optimizable is the calls to unicode_normalize.