We tried it a few weeks ago, and it did not work on all supported architectures (maybe a kernel compatibility issue).
So, the benefit of using zstd for only a few architectures vs the added complexity of having different squashfs across the board did not convince us of pursuing that option for the moment.