PyPy seems prone to intermittent SQLite failures.  An immediate retry
avoids them.  Not great, but it works.

It was moved to __init__ to avoid recalculating, but the directory could
have changed, so we need to wait to do the work.  Instead, only do the
relpath on systems that need it (Windows Py 2).

Since the self.filename attribute doesn't change
during the lifetime of a SqliteDb object,
we can move its relpath transformation in
the init method, instead of doing it every time
_connect is called, resulting in a ~30% performances
gain.