As far as I understand, this shouldn't happen. These connections should properly time-out like they do without flow offload. Because eventually, the conntrack table will fill completely and you won't be able to open any more new connections. I haven't actually tried that yet myself though. I turned off flowoffload when I noticed this bug.
hw flow offload on mt7621 also offloads PPPoE. I am seeing 97-98% idle CPU usage while fully loading my 500/500 mbit connection. I was simply trying to figure out how I can keep the PPPoE offloading enabled and do the shaping on a dummy ethernet interface. Haven't had time yet to try this myself, but @dlakelan has a very interesting post regarding this on his blog: http://models.street-artists.org/2017/12/11/inbound-qos-with-virtual-ethernet-and-policy-routing/
Amazingly cool stuff 