Adds functionality to train on TPU pods or slices

[jvrsgsty]

Now you can specify a tpu pod or slice in tpu_name, but will require an additional single device eval_tpu_name to be specified to perform evaluation on, as eval is currently unsupported on pods.

Also accounts for different locations of the objects required for TPUTraining:

• TPUEstimatorSpec
• TPUEstimator
• RunConfig
• TPUConfig
• CrossShardOptimizer
• TPUClusterResolver