Here you can find the original proposal for smart pointers: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2003/n1450.html
It answers your question quite precisely:
Since the deleter is not part of the type, changing the allocation strategy does not break source or binary compatibility, and does not require a client recompilation.
This is also useful because gives the clients of std::shared_ptr some more flexibility, for example shared_ptr instances with different deleters can be stored in the same container.
Also, because the shared_ptr implementations needs a shared memory block anyhow (for storing the reference count) and because there alreay has to be some overhead compared to raw pointers, adding a type-erased deleter is not much of a big deal here.
unique_ptr on the other hand are inteded to have no overhead at all and every instance has to embed its deleter, so making it a part of the type is the natural thing to do.
