RFC: Drop support for full TensorFlow runtime on macOS

This issue is for discussing dropping support for the full TensorFlow runtime on macOS. One recent issue where the difficulty of supporting full TensorFlow on macOS has come up is #3317, but there are numerous cases. We have to manage the macOS CI workers ourselves and the full TF runtime build is super resource intensive, which causes lots of downstream problems: long waits for full task groups to complete and dependent tasks (such as NodeJS packages + tests) to run, always needing a bit more worker compute power for macOS and not having it, dealing with storage space leaks and other macOS-only problems and needing to waste time fixing those instead of fixing bugs in DeepSpeech itself or reviewing PRs, adding features.

The way I see it the full TensorFlow runtime is only useful for using DeepSpeech models on the server-side. For example, by using TensorFlow serving tools. For that use case, macOS support is irrelevant. This issue is to propose dropping support for the full TensorFlow runtime on macOS and leaving only TensorFlow Lite support.

If you depend on the full TF runtime on macOS and have a good reason to do so, please flag it here so we can discuss. If we get no replies in one month, I'll move forward with the work to drop support.