How can you improve the performance of write workloads in a multi-regional Cloud Spanner instance that are significantly slower compared to a regional instance for a write-heavy application?