Raise error on training data from different format versions

[YodaEmbedding]

Perhaps an error could be raised to warn if training data of different formats is mixed? It took me a bit to figure out why it was failing until I realized that the record sizes being read were varying.

---

EDIT: Hmm, using any non-v6 training data runs into problems since parse has v6_gen hardcoded into it:

    def parse(self):
        """
        Read data from child workers and yield batches of unpacked records
        """
        gen = self.v6_gen()        # read from workers
        gen = self.tuple_gen(gen)  # convert v6->tuple
        gen = self.batch_gen(gen)  # assemble into batches
        for b in gen:
            yield b

[teck45]

T60 since march 15 is 100 percent v6, 2 weeks before mostly v6 and around 10 percent v5 - needs rescoring.. t75 is 100 percent v6. Also some data can be corrupted. But overall I have good experience with late t60 data.