[Draft] XLA and AMP

[ekuznetsov139]

• This PR adds support of XLA. Specify a command-line flag --use_xla=1 or --use_xla=2 to enable. With --use_xla=1, XLA will be used to fuse several specific subgraphs like AdamWeightDecayOptimizer. With --use_xla=2, TF will try to fuse the entire graph to the maximum extent possible.
• It enables AMP via the flag --use_fp16=True (supersedes the branch https://github.com/ROCmSoftwarePlatform/bert/tree/enable_AMP).
• Alternately, it enables fp16 via the flag --manual_fp16=True. This code was lifted straight from NVBERT and has not been tested.
• It adds continuous logging, letting you see the loss in realtime (also lifted from NVBERT).
• It adjusts evaluation logic.

It's been tested with horovod+xla+adam with and without fp16 and it seems to work correctly. With 8x MI50, seq 128, batch size 10, 1M steps (125K/GPU), final loss is 2.179 +/- 0.003 with fp32 and 2.202 with fp16.

It may be necessary to get a very recent build of TF for horovod & xla to work together.