Multiple files on bot tranning

[MatheusFaria]

This PR closes issue #21, if you pass a folder to the HUBOT_CORPUS env var it will use all YAML on that folder. If you pass a file, it uses just that one file.

It also removes the need for the training data to be inside the hubot-natrual project. It can be any path now, the default is traning_data/corpus.yml.

A small fix on the Dockerfile was required to work, some dependencies, like git gcc python g++ make, are needed during runtime. Because node-gyp rebuilds when hubot runs.

[CLAassistant]

All committers have signed the CLA.