Make behat scenario restart more fault tolerant

We seem to be getting some bad responses from Selenium periodically and thanks to the architecture of Behat we currently end the session in a hacky way and without giving any useful information.

We should:

1. give useful debugging information
2. only stop behat if we fail to get the session on the first attempt to get a session
3. fail a single test instead of all tests

This is a long-standing issue which has no doubt become more prominent thanks to MDL-69072 which restarts the session between every test.