acquire a global lock to init the OpenSSL singleton locks

[reaperhulk]

...then init the library since we really only want to do that once anyway

This approach was suggested in #2299. Of note, in a subinterpreter environment a "random" subinterpreter will register the static locks and engine. This means it will contain the Python callbacks for them. If the subinterpreter is subsequently destroyed, do the locks and random engine continue to function? If not, this solution will not work.

This should be considered WIP until we have a solid answer to the above.

[codecov-io]

Current coverage is 99.98%

Merging #2336 into master will not affect coverage as of db77358

@@            master   #2336   diff @@
======================================
  Files          118     118       
  Stmts        12133   12139     +6
  Branches      1336    1336       
  Methods          0       0       
======================================
+ Hit          12131   12137     +6
  Partial          2       2       
  Missed           0       0       

Review entire Coverage Diff as of db77358

Powered by Codecov. Updated on successful CI builds.

[alex]

This should definitely be inside of a finally block. I still need to think of any of this makes sense though :-)

[reaperhulk]

Since this doesn't work I'm closing it. We'll do what's in #2446 probably but a complete fix is still elusive.