Revisit the design of get_matrix3d

[furgalep]

I've noticed that, for every eigen type, one has to reimplement class get_matrix3d. This seems error prone and tedious. For example, if we want to support transformation of 4x1 types (see #30), we will have to also implement class get_matrix4d in every type. Then if we want to support native dynamic matrices (see #29), we would have to implement something similar in every derived rotation type as well.

Can we find a simpler way that only needs one implementation per linear algebra package?