Parametrize SparseMatrixCOO with vector types

Currently, a SparseMatrixCOO can only store explicit vectors. It would be convenient to define `rows`, `cols` and `vals` as `view`s, for example in a factorization-based Levenberg-Marquardt solver, where we will factorize the augmented Jacobian

$$
\begin{bmatrix}
J \\
\sigma I
\end{bmatrix}
$$

but we want to update $J$ and $\sigma$ separately.

I think that can be done by parameterizing the `SparseMatrixCOO` type with the "vector" types.

@farhadrclass