You should probably submit a bug report for this.
However, if you look inside the spec for mlmodel files you can see that it supports only a limited number of very specific models. But TensorFlow lets you create totally wild computational graphs that don't really fit those fixed model definitions. So if your TF model is compatible with the things supported by mlmodel, you can convert between the two. But in that case it's probably easier to write your own conversion script.
You don't have to. PerfectlySoft just released Perfect-TensorFlow, an open source Server Side Swift library for TensorFlow support, which can import TensorFlow model natively. The documents and examples are coming soon.
This just runs Tensorflow, which you can already do with their C and C++ APIs (and which I am currently doing and running inference on). I dont see how Perfect-Tensorflow applies to CoreML.
However, more importantly, I forgot that Keras can use Tensorflow as a backend, and Keras is supported in CoreMLTools (albeit, not 2.0)