I wonder if Tesla enabled the map generation with a new firmware update... however, it would only really explain increased drain with AP2 hardware though, so it's probably not the case here... something to think about though.
Essentially, mapping is one of the unmentioned pillars of FSD that Tesla never talk about. I would imagine they're going with Nvidia's solution, which is part of the capabilities of Driveworks/DrivePX2...
Instead of sending out loads of specially equipped 3D lidar style mapping vehicles, they crowd source HD maps from 'harvesters' - any vehicle with a camera. They start with pretty decent resolution 2D maps (most likely from HERE maps), then refine or build HD maps, using
motion and localisation data collected from the cameras.
End-to-end HD Mapping for Self-Driving Cars | NVIDIA
Obviously, you can't send back raw video to the data centres, so it actually crunches and pre-processes a LOT of information on the harvester itself - in Tesla's case, the DrivePX - before uploading the relevant information in a very compressed form to the data center. It really only does this intensive pre-processing when it's not actively driving - for obvious reasons. So, in Tesla's case, I would imagine that whilst the car isn't moving, it's crunching data it's collected from your recent trips. It may also explain how the car can seemingly get a bit 'better' on AP even without FW revisions - the HD map tiles it pulls down every 10km have been generated, or are more accurate.
tl;dr - your car is creating HD maps whilst you're not driving it.