Since the release of GRBD 7, Task cancellation interrupts asynchronous database accesses.

This has created a problem with transaction observers. If a Task is cancelled after a transaction was successfully committed to disk, observers could fail processing the transaction. ValueObservation, in particular, could not fetch fresh values, and report an error:

1. A task is started and performs an asynchronous write.
2. The task is cancelled after the write was committed to disk, but before observations could handle the change.
3. Observations attempts to fetch a fresh value, fail, and report an error (that's a bug).

This pull request restores the independence of transaction observers, and makes sure they can access the database even from a cancelled Task.

Thank you @tgrapperon for reporting this issue in #1746