What are valid usages? Why don't ISPs prevent this by default? Do any major ISPs do this already?
The RFC proposes that ISPs block traffic coming into their network with source IPs that are not registered to come from the devices passing the traffic along. There are a lot of details to this, but in short there's a protocol known as Border Gateway Protocol (BGP) which has routers on the edge of an ISPs network broadcast which IP ranges they can route traffic for. When one ISP connects their network to another ISP (known as peering) they share this routing information. Because this information is known and shared, it's easy to detect when traffic contains spoofed IPs because that connected network doesn't advertise that they route traffic for those IPs, hence it can be blocked.
The problem is that ISPs don't generally implement this because (they claim) it costs them too much in resources to perform the extra analysis of the packets to determine whether it's spoofed. This is why it's still possible to spoof IPs. I believe there are some ISPs that do implement but not all do.
Even if ISPs implemented it, you can still spoof IPs as long as they are in the proper IP ranges for which the ISP broadcasts that it can route for. While it's more limited, that range can still be quite large.
In addition to the technical answer Michael has provided, I'd provide an economic one.
In short, egress filtering is what economists call a "negative externality". Briefly, a negative externality is where the cost goes to a different entity than the gain. In this case the cost is to the ISPs, and provides little or nothing to them in return. The benefit happens to the internet at large, and only happens when almost every ISP in the world does it.
It's worth mentioning again that the benefit would only be gained if a very large portion of ISPs implemented this. The global nature of the internet makes this very unlikely. If implementation isn't nearly universal, then attackers only need find one ISP that doesn't do egress filtering.