Probe embedding geometry

The whole probe is one class, Embeddings, which normalises every vector once so that all similarity is a single matrix-vector product. Around it sit a download script, a short exploration script and the experiment, which asks every a : b :: c : ? question two ways. The decision that matters most is scoring both ways on every question and also counting how often the top hit is one of the inputs: that third column explains the gap between the two accuracy numbers.