Investigate RTLD_DEEPBIND dlopen flag

[diegoferigo]

In the past, we faced countless problem in the interaction of tensorflow installed from PyPI and ScenarIO. Also the need to use singletons introduced some challenges since the very beginning of the development.

As workaround, over time we implemented the following:

• (old) Import tensorflow first
• Automatically preload tensorflow library (Force loading tensorflow shared libraries to prevent segfault #290)
• Automatically preload gym (Fix segfault with protobuf, gym (atari), cv2, and ScenarIO #316)
• Alter dlopen flags to make singletons work (Ignition robotics libraries and python: plugins symbol resolution problems #7)

I hope these are all the workarounds, if I remember other ones I'll link them here.

Today @traversaro noted the RTLD_DEEPBIND dlopen flag that could be worth trying. I might have already tried it in the early stage (#7), but if I did, results were inconclusive and I didn't log them. Now that we have more knowledge of the overall problem, maybe we can give it another try.