gracefully handle incompatible hardware / handle server crashes

[k-dominik]

I just stumbled over an issue on apple M1 - our current tensorflow version (1.14) doesn't run via rosetta - it fails with Illegal instruction... (the python process crashes/exits). This whole thing is somehow hidden in tiktorch - when, at some point tensorflow is imported in the server, the server process simply disappears.

For known incompatibilities like tensorflow with M1, we can of course issue an appropriate message before importing at all. Things like this might also happen with other libs - for this we should better handle crashing of the server.