Fix multi-gpu metrics report

[ymwangg]

This PR fixed #2508. I'm not sure why GPU device does not allow reusing the same ClusterSpec but leaving it blank seems to work.

Tested with GPU_NUM_DEVICES=8 python test_train_mp_mnist.py --metrics_debug --fake_data --num_epochs 1.

Single-gpu metrics report: https://gist.github.com/ymwangg/609b79ed783a1797f87d55480d23bc8d
Multi-gpu metrics report: https://gist.github.com/ymwangg/8218ddf4ed8ce2d1ac35758effabdf6f

cc @YangFei1990

[YangFei1990]

Great, thanks a lot @ymwangg! A few questions for my understanding:
If we do not use the cached session config will it generate some default config? How are we sure that such change won't affect the metrics generation?

[ymwangg]

@YangFei1990 My understanding is that the same graph will produce the same results regardless of different session configs. It could be wrong and Google folks may have a better answer.

[JackCaoG]

Thanks! I think rebase should fix the build issue.

[JackCaoG]

Thanks @ymwangg !