Fixes, fixes, fixes

• Update to ndarray 0.12
• ff-train: show a progress bar while reading the vocabulary.
• Fix a bug in negative sampling for the structured skip-gram model.
• Remove dependency between cloned RNGs.
• Fix one-off in the loss computation, resulting in reporting losses that were too low.
• Add ff-compute-accuracy for evaluating anologies.
• Do not unnecessarily sort the vocabulary upon deserialization.
• ff-train: add the zipf option to specify the exponent of the zipf distribution.

Memory-mapped models

• Support memory-mapped embedding matrices. This makes loading of the embedding matrices instantaneous and reduces memory use. The use of memory-mapped matrices comes at a small cost in efficiency.
• Normalize stored embeddings with their l2 norms. This avoids normalization when loading a model and simplifies the functionality for similarity/analogy queries. The l2 norms are stored, so the word vectors can be restored with their original magnitudes by multiplying them by their l2 norms.
• Add subword representation embeddings to known words in stored models. This speeds up the retrieval of known word embeddings. There is no loss of information: the original word embeddings can be reconstructed by subtracting their subword embeddings.