1) Basically with Amazon ELB you can't predict where a session is going to landing but the logic of the Apache httpd load-balancer will have to connect to the right node.
So for one request you may need up to 3 connections to the same tomcat node. So you need to configure tomcat to handle a lot of connections.
2) 200 in httpd and 600 in Tomcat :-(
3) the httpd should get the same load so at some point the whole system will hang if you don't define the max correctly (depend on the webapp).
You can use the NIO or APR connector in Tomcat so you can have a lo of connections and not that many threads.
your info (in particular number 2) is really useful. Actually i was following the suggestion coming from http://lbconfig.appspot.com/ that doesn't point out the right number of maxClients/maxThreads.