All exceptions, if not caught, are unconditionally sent to `sys.stderr`.
We only need to use our own custom class with some modifications and we
can intercept all errors, even those in threads (which #151 was about).
The only real downside from this practice is that `cli` is not
accessible to us when the errors happen, so I used a hack to set it on
an instance variable. If anyone can find a better solution, please step
forward. For the time being, this will have to do.
If for some obscure reason this class breaks, `sys.__stderr__` holds the
original `sys.stderr`
