Fast averaged-perceptron binary classifier.  Includes weighted-tf feature
extractor.  Weight and feature vectors represented as cache-friendly
std::vectors, which resulted in a 4-5x speedup over the previous
representation as std::maps.

Evaluation:

Initial rough-cut accuracy of 95.8% achieved with 2 training
iterations on the news20.binary dataset at
http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary.html.

Training set consists of 8000 examples: 4000 of +1 and 4000 of -1
taken from news20.binary.  Remaining 11996 examples used for testing.

Thanks to Justin Palmer for explaining the difference between an
averaged and a regular perceptron as well as teaching me the lazy
update 'trick'.  It caused a huge speedup.

Any mistakes in the code are mine alone.