Cover Savepoint rollback-to scope-exit semantics

[SRombauts]

Summary

• document and test that ROLLBACK TO leaves a savepoint active
• verify that writes made after a manual rollbackTo() are still rolled back on scope exit
• verify that writes made after rollbackTo() can still be explicitly committed with release()
• clarify the destructor's catch-all cleanup comment

Rationale

A deep review of the mbRolledBack member introduced by #559 and removed by #576 found that it should not be restored.

SQLite's ROLLBACK TO rewinds to the savepoint but keeps that savepoint active. If mbRolledBack permanently caused the destructor to skip its final ROLLBACK TO, then changes made after a manual rollback would be committed by the destructor's RELEASE instead of rolled back.

The existing implementation—always ROLLBACK TO before RELEASE on scope exit—is therefore the correct RAII behavior. This PR locks that behavior in with regression coverage.

No ABI or public API change is made.

[coveralls]

coverage: 99.715%. remained the same — savepoint-rollback-semantics into master