We can log for different threads (main, map, workers, etc) so
we need to lock before adding the messages on the list.
This was causing few tests to randomly crash.
- catch SQLite exceptions and report them
- failed statements are ignored, we're really just caching here, so if it fails we're handling it gracefully elsewhere
- handle cases where the database file goes away after we opened it
- handle cases where the schema wasn't created after the database file was opened successfully
- add tests
Now the FixtureLog is an Log::Observer instead of a backend. The
observer can sniff and filter log messages, that can be used
for fail a test if it emits a warning for example.
