You clearly have 5 sites there, so that eliminates option 1. Sites don't necessarily have to correspond to geography, but in reality it practically always works out that way.
There is no connectivity between several of the sites. I.e., the network is not fully meshed. That means you have to unbridge all site links and create specific site links, so that immediately eliminates option 2.
Just remember that even if you have 20 sites that are all fully-routed/meshed, and you have one satellite site that can only talk to one of the other sites... then you must unbridge all site links on account of that one remote site.
Sites that are connected should have their site links bridged.
So you will have:
1 site link New York <> Atlanta.
1 site link Sydney AUS <> Atlanta.
And since Washington, Boston and Atlanta are all fully meshed with equal bandwidth, you can make 1 site link contain all three of those sites so that replication can occur transitively among those sites.
So the answer is option 4. Five sites and three site links.
Option 3, creating five site links, an extra one for Boston <> Atlanta, Washnigton <> Atlanta, and Washington <> Boston, would work OK, but by doing that, you forfeit the transitivity since all sites contained within the same site can replicate transitively. Plus, why create more site links than you have to? It's difficult to really drive that point home in this scenario, but would become a more important factor if your network ever expanded to have more sites.
To further elaborate, just for the sake of completeness - there is a situation where you might want a more extravagant solution for the Washington-Atlanta-Boston network than just a single site link with all three sites in it. And that is if you were bandwidth-constrained between two of the sites - you might want to make a dedicated site link to represent that connection - but give it a higher replication cost, so that the KCC will prefer to replicate across the higher bandwidth, lower cost link... then use a site link bridge to bridge all of the Washington-Atlanta-Boston site links, so that AD would know that it could replicate transitively across the other link in case one of the links went down, but during times of normal operation it knows to prefer the better link. Remember that when calculating site link costing, both site link costs in the transitive route added together need to beat the site link cost of the single point-to-point link, if you want the longer route to be preferred.
But in the words of Einstein, it's best to make things as simple as possible, but no simpler.