Also removed native binaries being packed with TensorFlowSharp. Not only do they make the package unreasonably large due to cross-product of native runtimes, but they also remove the flexibility of specifying which exact native TF runtime is used for inference. This is often required in practice since the TF version used for inference should match the TF version used in learning. Later it would be possible to distribute individual runtime packs for specific TF builds to bring back the plug-and-play capabilities.
Fixes #479
Also removed native binaries being packed with TensorFlowSharp. Not only do they make the package unreasonably large due to cross-product of native runtimes, but they also remove the flexibility of specifying which exact native TF runtime is used for inference. This is often required in practice since the TF version used for inference should match the TF version used in learning. Later it would be possible to distribute individual runtime packs for specific TF builds to bring back the plug-and-play capabilities.