Large model evaluation

[naveenram00]

Right now the model evaluation script feeds in the checkpoints one at a time based on the range [FLAGS.eval_start, 999901 + FLAGS.Steps]. This works for the base size because the base size model starts at 999900 steps for finetuning. This starting value is different for some of the other sizes of the model so the code needs to be modified to run evaluation over the right range for those models.