python: consider a backtrace handler for debugging

[lidavidm]

What feature or improvement would you like to see?

We've gotten occasional reports of things crashing without further explanation that have been hard for us to reproduce (#1531, #1522, #1476). In these cases, being able to at least get a stack trace would be useful. I would propose the following:

• Statically link backward-cpp or similar in the Python driver manager with enough dependencies to produce a full backtrace
• Enable the backtrace handler on startup via an environment variable
• Build Cython code and all drivers with debug info so that we can get a backtrace
• Include backward-cpp in the LICENSE since we would be shipping it statically

TODOs:

• This may increase binary size a lot, should we consider separate packages? (Does backward-cpp work with something like debuglink? It seems yes: https://github.com/bombela/backward-cpp/issues/160. At least for wheels, we could either ship a separate wheel in the GitHub release with debug info, or we could split off the debug info and offer it as a separate download. We should measure this.)
• How should we get Golang backtraces into this?

[lidavidm]

We can't do this because the default backtrace mechanism only recognizes exported symbols (making the resulting backtrace not very useful) and the more advanced backtrace mechanisms require linking (L)GPL libraries (which we obviously cannot do). We can at least build with debug info and let people figure out backtraces themselves.