[WIP] Matrix abstraction

[dance858]

uncommented windows, remember to add later

[Transurgeon]

really nice first iteration at cleaning this up with the matrix class.
One thing that could be added for this PR would be a constructor for right matmul (dense), what do you think?
The speedups with the batch matrix-vector products can be added later.

[Transurgeon]

claude suggests that we could use batched matrix-vector (with dgemm) products with a batch size of 256.
This seems to enable the major speedups. But we can do that in another PR.

[dance858]

Have you actually tried to see that it's significantly faster? It is certainly possible to do so, but it would probably require us to collect blocks of J first which likely would lead to much more complicated code.

[Transurgeon]

yeah I did, it makes a bit difference. but let's revisit this in another PR.

[dance858]

Awesome! Let's do that in a new PR.

[Transurgeon]

claude suggests we could do this to avoid dynamic allocation with iVec:

 The sparsity function uses iVec (a growable array) to accumulate row indices. Since dense
  A means every non-empty block contributes exactly m rows, we know the nnz per column
 upfront. A two-pass approach (count, then fill) avoids dynamic reallocation:

 // Pass 1: count nnz per column
 for (j = 0; j < k; j++) {
     col_nnz = 0;
     for each block with entries: col_nnz += m;
     Cp[j+1] = Cp[j] + col_nnz;
 }
 // Pass 2: fill row indices directly into C->i

[Transurgeon]

I think we can merge this!
There seems to be some code formatting issues though.