Robustness fixes in test suite machinery. Added a module-level INSTALLED flag, which can be set to false if the test suite is being run in-place (without ipython having been installed at all). This is because how we call and import things must be done differently depending on whether the code is installed or is being run in-place. The only ones that can know this reliably are the entry-point scripts, so those are responsible for setting this flag. Also made the code that validates ipython in subprocesses report errors better, by checking stderr for errors before validating stdout output, as anything on stderr will be likely informative of the real problem.