Anyone?
I think the best place for the pppoe server is centrally-located.
You could even put in a redundant pair at the central site for more reliability.
Furthermore, I would use layer 3 in your backhaul topology. It's going to be more scalable and flexible for you to roll out a layer 3 topology with dynamic routing. VPLS will give you the "bridge to anywhere you want" ability, so there's no need to bridge across a bridge. That just invites trouble. MPLS is designed to ride on top of an intelligent layer 3 topology, so use that to your advantage.
You can add extra point-to-point shots directly between remote towers for extra redundancy, and ospf/MPLS will use them as backup routes automatically. (suppose you want to add a link from the 1:105 / 1:106 site back to the main site directly - this is easy with OSPF and much faster to reconverge in case of failure than STP would be.)
So now the APs, aggregation routers (in stead of L2 switches), and central MT router make a dynamic topology. Use VPLS between the APs and central router (as the diagram shows), and then apply the PPPoE server to the VPLS bridge at the main site. This is where you put TX limit on the user profiles. It's clean, easily managed by profiles, and limits the bandwidth before it crosses your backhaul infrastructure.
EDIT: I was thinking TX downstream (sorry, to me, every interface is relative to the local router). For what you're asking, I would put TX limit in each CPE SXT and set it to the largest upstream bandwidth class that you sell. This will keep users from running away with upstream bandwidth before it crosses your backhaul only to get dropped at PPPoE server.