Fix for the broken "--config=rocm" build (followup to PR 20277)

[deven-amd]

This is more of a followup to PR #20277

As per the following comment by @yifeif
#20277 (comment)

if_cuda_is_configured was changed to if_cuda in the merge to make some internal targets pass. Unfortunately it seems that doing so break the ROCm build.

When I try to do a ROCm build on a clean repo (post PR 20277 master branch), I get the following errors

  Successfully uninstalled tensorflow-1.10.0
ERROR: /root/tensorflow/tensorflow/contrib/tensor_forest/BUILD:110:1: Label '//tensorflow/core:gpu_lib' is duplicated in the 'deps' attribute of rule 'tensor_forest_kernels'
ERROR: /root/tensorflow/tensorflow/contrib/tensor_forest/BUILD:220:1: Label '//tensorflow/core:gpu_lib' is duplicated in the 'deps' attribute of rule 'model_ops_kernels'
ERROR: /root/tensorflow/tensorflow/contrib/tensor_forest/BUILD:315:1: Label '//tensorflow/core:gpu_lib' is duplicated in the 'deps' attribute of rule 'stats_ops_kernels'
ERROR: /root/tensorflow/tensorflow/contrib/boosted_trees/BUILD:324:1: Label '//tensorflow/core:gpu_lib' is duplicated in the 'deps' attribute of rule 'model_ops_kernels'
ERROR: /root/tensorflow/tensorflow/contrib/boosted_trees/BUILD:395:1: Label '//tensorflow/core:gpu_lib' is duplicated in the 'deps' attribute of rule 'split_handler_ops_kernels'
ERROR: /root/tensorflow/tensorflow/contrib/boosted_trees/BUILD:439:1: Label '//tensorflow/core:gpu_lib' is duplicated in the 'deps' attribute of rule 'training_ops_kernels'
ERROR: /root/tensorflow/tensorflow/contrib/boosted_trees/BUILD:486:1: Label '//tensorflow/core:gpu_lib' is duplicated in the 'deps' attribute of rule 'prediction_ops_kernels'
ERROR: /root/tensorflow/tensorflow/contrib/boosted_trees/BUILD:534:1: Label '//tensorflow/core:gpu_lib' is duplicated in the 'deps' attribute of rule 'quantile_ops_kernels'
ERROR: /root/tensorflow/tensorflow/contrib/boosted_trees/BUILD:578:1: Label '//tensorflow/core:gpu_lib' is duplicated in the 'deps' attribute of rule 'stats_accumulator_ops_kernels'

The following commands can be used to do the ROCm build

<run configure.py to enable the ROCm support>
bazel build --config=opt --config=rocm //tensorflow/tools/pip_package:build_pip_package 

This build does complete without any errors with the commit in this PR.
I realize that this commit will most likely cause regressions in the internal targets/builds that @yifeif eluded to in the comment. So we need to come up with a solution that makes everything work, and we will need your help in that matter.

[dagamayank]

/cc @gunan @whchung

[dagamayank]

@yifeif did you get a chance to review this PR?

[yifeif]

I just submitted ea42e88, which should make the macros behave as what #20277 originally intended. Let me know if the change at head works for you, and we can close this PR and move forward. Thanks for the patience. Sorry it took a bit longer to get this work both internally and externally.

[whchung]

closing this PR based on @yifeif 's statement