Untiful sparse arrays in matrix equation solvers

Hello :-)

Does anyone know a non-allocating way to create a unit-stripped view into a `SparseMatrixCSC`? Let's assume I have the following sparse array:

```julia
A = (sprand(3, 3, 0.5))u"1/s"
```

Now if I want to use `ustrip` directly on that matrix, it'll default to `ustrip.(A)`, which will allocate memory (which we want to avoid). A workaround could be using a reinterpretation of `A`, as it is done in the `ustrip()` implementation for the `Matrix` type:

```julia
reinterpret(Unitful.numtype(eltype(A)), A)
ERROR: `reinterpret` on sparse arrays is discontinued.
Try reinterpreting the value itself instead.
```
:-(

Now let's assume we have the following idea:

```julia
nzval_ustripped = ustrip(A.nzval)
A_new = SparseMatrixCSC(A.m, A.n, A.colptr, A.rowval, nzval_ustripped)
ERROR: MethodError: no method matching SparseMatrixCSC(::Int64, ::Int64, ::Vector{Int64}, ::Vector{Int64}, ::Base.ReinterpretArray{Float64, 1, Quantity{Float64, 𝐓 ^-1, Unitful.FreeUnits{(s^-1,), 𝐓 ^-1, nothing}}, Vector{Quantity{Float64, 𝐓 ^-1, Unitful.FreeUnits{(s^-1,), 𝐓 ^-1, nothing}}}, false})
```

The problem here is that `Base.ReinterpretArray` is not a `Vector`, which is required by the `SparseMatrixCSC` type. And now I'm out of ideas ;-)

Does anyone else have an idea how we can have unit-stripped views into sparse matrices which can then be used for efficient linear algebra?

Start by examining Unitful's ustrip implementation for Matrix and the SparseMatrixCSC constructor behavior described in the issue. Investigate how A.nzval and reinterpretation interact with sparse arrays. Done means providing a non-allocating unit-stripped view that can be passed to efficient linear algebra routines.

Written by the indexing model from the issue text.