Build: Remove WORKSPACE dependency on TensorFlow

As discussed during the SIG-TensorBoard meeting, heres an issue with more detail. @nfelt

Currently TensorBoard has an http_archive rule in the WORKSPACE file depending on tf. It's been around for a long time and hopefully we can migrate away from it soon.
Eg: https://github.com/tensorflow/tensorboard/blob/cc01bc1eb57d123511dc0d8a659cb19a9c979976/WORKSPACE#L66-L74

Most other tensorflow-related packages are now based off how https://github.com/tensorflow/custom-op works, if possible TensorBoard should do the same. That way it'll build against the exact version installed (and the exact version that distro packagers ship) Uncoupling the exact version of tensorflow would also make the whole ecosystem easier to build and CI can build against TF master more frequently.

I realize that removing it may take a while, and from the SIG-Build side, I'd love to find out more about exactly what parts of the TF tarball you need to build. I want SIG-Build to put together a common set of build rules that all these related packages can use so if tensorboard needs things that custom-op does not yet provide then i'd love to hear more.

Start with the TensorBoard WORKSPACE file, especially the http_archive rule linked in the issue, and compare it with the tensorflow/custom-op build approach. Identify which parts of the TensorFlow tarball TensorBoard needs for builds. Done means removing the WORKSPACE dependency while preserving builds against the exact installed TensorFlow version.