Build fully automated build chain for TensorFlow and all other kernels

[achimnol]

TensorFlow v1.7 will be the last version that supports our current CUDA 8.0 + cuDNN 6.0 build chain.
From TensorFlow v1.8 we need to upgrade CUDA.

Currently we build the images on a dedicated physical machine, which has a single CUDA version.
For maximum stability and automation, it would be nice to run our builds on spot p2.xlarge/p3.xlarge instances with an appropriate Amazon DeepLearning Base AMI (v4.0 or v6.0).

Let's write scripts to do this.

• Create a cloud build configuration that supports:
  • Automatically trigger the build process by git pushes to this repository and the kernel runner releases on PyPI
  • Build only modified Dockerfiles but with dependency checks to base images
  • Ability to force-rebuild specific images (manual trigger)
  • Push rebuilt images to the docker hub and designated private docker registries (for enterprise customers)
• Optional but good to have
  • Save/load tarball'ed docker images for cache heating for docker builds (maybe from/to S3, or utilize EFS) -> comparison test required
  • Automatically run basic code execution tests against newly built images
    • maybe using ansible, pupeet, vagrant, etc. on temporary p2/p3 instances

[inureyes]

@hephaex Please check this issue.