Simplified tensor building

Make it simpler to create sparse labeled tensors. Two approaches are currently in use: nested lists in the tests, and explicit item assignment in AnalogySpace and other code like that.

I'm thinking about making tensor builder classes that make a tensor and updates it with a passed sequence, which will probably be a generator. Then `queryset_to_tensor` would look something like:

    def queryset_items(queryset):
        return (((concept1, relation, concept2), get_value(score, freq))
                for (relation, concept1, concept2, score, freq) in queryset.values_list(
                'relation__name', 'concept1__text', 'concept2__text', 'score', 'frequency__value').iterator())

And you'd call it like:

    Mirrored3DTensorBuilder.from_seq(
        queryset_items(conceptnet_queryset(cutoff)),
        identity_weight)