Re-think TensorFlow / DeepSpeech build dance

[lissyx]

Now that we masterize better the build-and-cache aspect of TaskCluster, it might be a good time to re-think the interfacing for the known-to-be-painful build dance between TensorFlow and TaskCluster:

• [x] make TensorFlow a git submodule #3107
• [x] only build libdeepspeech (tensorflow and tflite) #3107
• [x] move tensorflow taskcluster bits into deepspeech repo: one place to build them all #3107
• [x] leverage bazel-level caching for build artifacts #3107
• [x] rework the current taskcluster build-or-cache templates to be more generic and use them for tensorflow? (or copy/pasta) #3068
• [x] re-do macOS workers setup #3284 #3289
• [ ] interface bazel build with make ?

[imskr]

Can I help on this @lissyx

[lissyx]

Can I help on this @lissyx

Well maybe not the TC things, but if you want to experiment smoother flow with tensorflow as a submodule, you're welcome

[imskr]

Cool. Which branch we have to make git submodule? @lissyx

[lissyx]

So I have a PR with:

• tensorflow as a submodule
• taskcluster scripts / definitions from tensorflow moved into deepspeech
• task graph dependencies between deepspeech and tensorflow
• definitions of tensorflow builds that should be using the generic build/cache mechanism

still a lot to complete, but it's on track.

[lissyx]

* [ ]  interface `bazel build` with `make` ?

@reuben Maybe GitHub Action move is a good time to address this?