Store patterns in BitMatrices instead of sets

[gdalle]

I want to try this but it might require larger code changes than I thought.
A first hurdle is myempty: for a set it doesn't require further information, but for a sparse array we need to give the ambient size.

[adrhill]

I just read your comment in #135 (comment).
Let me retry to add pattern types first, then we go back to this and #135.

[adrhill]

As mentioned in #135 (comment), depending on the problem, it might actually be reasonable to use "dense" BitVectors and BitMatrices if we just allocate them once.

[adrhill]

Do you still want this or can I close the issue @gdalle?

I can't imagine it being faster than DictHessianPattern since SparseVector and SparseMatrixCSC are based on Vectors of indices and values.

[gdalle]

I think for small and medium problems using a BitMatrix to store the sparsity pattern in shared (Walther) hessian mode would crush the competition

[gdalle]

So yeah I'd like to keep this issue open at the least