Let's assume you have two links to a destination and that you want to load-balance your traffic. And let's assume again that you are sitting at your PC at this end of this arrangement and there are multiple servers at the other end you might wish to connect to.
If you are using per-destination load balancing then when you open up a connection to remote server 1 from your PC the traffic for that server, all of it, will use the first link. Now you open up another connection to remote server 2 and that traffic will use the second link. Traffic for remote server 3 will use the first link again (adding to the traffic already running on there for remote server 1) and so on.
Round-Robin is effectively the same thing as this.
Now if instead you are using per-packet load balancing then the individual packets you send from your PC are shared out across the two connections in the order that you transmit them and regardless of their destination. Thus you can have packets for remote server 1 for example going over both links rather than just one as before.
So which would you choose? The issue to think about here is whether there might be any difference in the two paths. If not then per-packet load balancing would probably provide a more accurate half-and-half division of the traffic between the two links. But if there is a possibility, say, of a topology change redirecting one of the two paths via a completely different route then you might end up with packets arriving out of order and requiring reassembly, poor application performance or even an asymmetric routing situation. In that case per-destination might be a more prudent choice even though it wouldn't divide the traffic as equally between the two links.
Hope that makes sense.