errors in Journal.prototype.rollback obscure actual errors.

[stratoukos]

When an error occurs in a transaction, Lovefield will attempt to rollback that tranasaction. If an error occurs while it is doing so, only the second error will be reported and the original error will be silenced.

Here is a minimal repro: https://gist.github.com/stratoukos/9ecf84a95b74413f708a56a640e06bb9.

As explained in #153, the above throws a 203 error. While the transaction is being rolled back, a 201 error occurs and the original 203 error is silenced. If you drop a breakpoint at the spot where the 201 error is thrown (BTreeNode_.prototype.insert), you can see on the stacktrace that this happens during the rollback because of a previous error.

[freshp86]

Lovefield should always be able to roll back a transaction without any problems. Not being able to cleanly rollback a transaction violates the atomicity aspect of the ACID database properties. So the real issue here is to figure out why the journal can't be rolled back.