I see, could you run top -d 1 during one of your tests tat shows ping spikes and look at the sirq percentage of the pi? For testing you could run a dslreports speedtest, or even the fast.com speedtest (that also allows to configure upload testing as well as longer run-times) while visually monitoring sirq.
Or you could try the speedtest package under https://forum.openwrt.org/t/speedtest-new-package-to-measure-network-performance/24647 as that will also look at the CPU load during the test.
That points to a bug or overload somewhere, thanks to your measurements we have proof of the overhead being 40 bytes, and encapsulation being ATM so pretending this was not true will run the shaper at higher rates than you would think and especially with small packets, not accounting for ATM cell encapsulation will underestimate effective packet-size by 505 to 33% and that will make your shaper ineffective against bufferbloat if the packet size mix on your link contains to many small packets. That said, unless the "bug" is fixed I do not blame you for using the "none" option as the reason for running sqm is not being theoretically correct but being practically better.
I would just like to figure out why link-layer accounting does not seem to work for you...