Type erased closures are strictly less useful than subtype polymorphism + virtual functions (another form of type erasure) as the latter allows you to share data among an entire interface.
Closures are simpler, and therefore less cumbersome for 90% of the use cases. I prefer this compromise over the other.
Also, I'm not sure I get this "type erasure" business. In ML languages, there is no such erasure. I guess you need this erasure because the language doesn't support parametric polymorphism to begin with? I'm thinking about Java, which hacked generics after the fact, and maintained backward compatibility through type erasure.
In C++, there are few type erasures. Generally, when we want a function to receive a closure, we write it like this (minus dark corner I have missed):
