If you really want the best speed and performance but at the cost of storage efficiency then mirrors are the way to go.
I have a video and forum post on the topic here:
Yes, but probably not linearly for this specific workload. In ZFS, IOPS scale with the number of VDEVs, while throughput scales with the total number of disks. So it will be a faster overall with a 4th VDEV.
Changing the dataset record size to 1M would help as well. PBS chunks are 4MB. A 1MB record size allows ZFS to store each chunk across exactly 4 records. At 128K (default), your 3× RAIDZ1 VDEVs are doing a lot of small-block parity work. At 1MB, the writes become much more “sequential-like” and should boost performance a bit, especially when doing garbage collection.