1.1) you can't think in term of one Thread by user. Depending of machine configuration but you could load thousands of threads but it will not scale and lose a lot of time in Thread context switch. Better think NIO Netty like framework with few incoming and outcoming Thread pool executor that perform incoming messages under milli second execution.
1.2) You can simply scale by putting in front of you game server a loadbalancer component that can forward incoming player to right server according their load
1.3) NIO can handle thousands to to millions connection on a single box. Don't worry with this.
2.1) Manage your player sessions and 2.2) be away of EJB architecture. It will eat all your box power instead of allocating power to your game which is your goal.
2.3) HTTP can serve all clients but if you run realtime game i encourage to use binary socket and keep HTTP only for loadbalancing , login , stats and fallback when can't establish a socket connection.
2.4) Socket based server can handle hundred thousands incoming message per second. This is the property of low latency system
While it's very interesting to dive in building such system. What is your goal? Learn to build such system or succeed your game?
Many java multiplayer game server technology framework already exist. SmartFox Server, ElectroTank...
We have our own java high load Nuggeta multiplayer crossplatform java game server that addresses all points discussed above and much more. If you wanna try it it's free.
If your goal is to write a game server it's awesome venture that can takes long time but very exciting. If your goal is to succeed your game. Pick up among Java game server SDK already existing.
Hope my thought can help you.
Best
