I believe there's an issue to consider here. Tensorflow is not a Python module. It is a C++ library that happens to provide Python bindings. Bindings for other scripting languages (e.g. Go) could be introduced in the future. There is a C API and C++ API as well, though these are - presently - not as easy to use [1].
I'm wondering whether it would make more sense to have two packages: "tensorflow" and "tensorflow-git", where both python2 and python3 bindings are installed by default. A similar approach is taken in the "gnuradio" package for example: it also provides a C++ library and the python2 bindings are implicitly installed.