For a general answer, please see section 5 of this answer:
Some more specific responses:
1. Correct – no effect on the consensus algorithm.
2. The address generation is not deterministic (although that's an interesting idea for us to consider in future) so you would need to use dumpprivkey/importprivkey for each new address.
3. See 2.
4. No problem sending a transaction to multiple nodes simultaneously, so long as the transaction is really identical (i.e. the same txid). In a peer-to-peer network transactions propagate in multiple directions anyway.
5. There's no sequence ID in the MultiChain architecture, but there is a risk of double spending the same UTXO. So unless you manually manage unspent outputs you can't use this kind of setup for load balancing. But you can use it for failover. In other words, only create transactions through one node at a time.