At this moment i would like to conclude by saying there is a lot of knowleges required to optimize network stack and measure CPU load correctly.
This should be my handbook: https://blog.packagecloud.io/eng/2016/06/22/monitoring-tuning-linux-networking-stack-receiving-data/
I tried to play with every option explained in an article above. Many of them are not supported by the hardware, some of them have no effects, but at least this article has a great explanation on network stack tuning.
Finally i was able to reduce displayed CPU load (mpstat) up to 77-83% IDLE with full CPU / max clock during 957 MBit/s receive with jumbo frames enabled (mtu 9000 on both sides) which makes me happy.
After a lot of tests I guess here is nothing wrong with Jetson / L4T network stack itself. Only some HW/SF/knowleges limitation which is pretty obvious.
Anyway, thanks for everyone participated, happy to work with such great community!