Edit: Removed previous analysis, as it is incorrect. Setting aql_enable to 0 essentially makes the ieee80211_txq_airtime_check() method always returns true. This essentially make the ieee80211_next_txq() method just schedule the txqs to be transmitted from left to right of the red-black tree for every round of scheduling. This has the side effect to making the first node (i.e. txq) of the rb tree always eligible for transmission regardless of how much airtime was spent for the txq. Maybe this is the source of the high latency?
So it seems that I'm not the only one affected.
In any case, I've reverted my R7800 back to the old round-robin algo. So far with 3 days uptime without issue (touch wood.) Will monitor it for a bit longer.