so it's probably the huge volume of sequential IO causing some buffers / caches to fill up and in turn block the IO of KVM guests.
For sure, zfs cand not have large sequential IO. If you have non-mirror zpool, it is very unlikely to have something like zis. Also if the free space in the pool is let say under around 85 % the seek time it will be higher, so IOPs will be bad.
I can not say nothing about qcow2, but I see that the performance /IOPs is very bad compared with raw.
And like I said before, try to avoid the results with zfs send-recive, for example.