For the second point, you can easily extract the state components by adding extra lines to the matrices C and D. For example if you have the following system with a state x of dimension n and an input u of dimension m

\begin{align}
x'&=Ax+Bu,\\
y&=Cx+Du,
\end{align}

you can define \tilde C=[C;\mathbf{I}] and \tilde D=[D;\mathbf{0}] where \mathbf{I} is the n\times n identity matrix and \mathbf{0} the n\times m zero matrix. Then the LTI system (A,B,\tilde C,\tilde D) will have an output \tilde y whose last n components are those of x.

What do yo mean ? It depends what you want to do with x afterwards. It is better to use a LTI system block and output the whole state x (taking C=\mathbf{I} and D=\mathbf{0}), then you can do what you want. But rebuilding the LTI block with elementary block is overkill, unless you need to add perturbations (like in your diagram above).

There is no smart or not smart way, this is just a matter of taste.

If you have a few number of states, e.g. two, I would not use the scifunc block and directly compose the diagram for individual states. However, if you want to take advantage of the matrix algebra, then use this block. Here is an example for the system (a spring with damping increasing with time):