It's pretty simple. Nobody wants to do ML R&D in C++.
Tensorflow is a C++ library with python bindings. Pytorch has supported a C++ interface for some time now, yet virtually nobody uses C++ for ML R&D.
The relationship between Python and C/C++ is the inverse of the usual backend/wrapper cases. C++ is the replaceable part of the equation. It's a means to an end. It's just there because python isn't fast enough. Nobody would really care if some other high perf language took its place.
Speed is important, but C++ is even less suited for ML R&D.
I think readability is what made python a winner. I can quickly type down my idea like a pseudo code, I can easily skim through other people’s algos. In C++ even a simple algo with a 100 lines of pseudo code will balloon to thousands of lines in c++