Sparse sum(dim=...) and chained groupby on CSR-backed expressions #964

Describe the feature you'd like to see

Make LinearExpression.sum(dim=...) keep a CSR backing instead of densifying. Child of #756.

This is the largest remaining hole in the sparse path. _sum (linopy/expressions.py:1694, 2214-2242) builds the dense rectangle and then stacks the summed dims into _term, so peak nterm becomes nterm x prod(summed sizes). After *, sum is the most common operation in model building, so a CSR-backed expression rarely survives to the constraint.

The same applies to a second groupby on an already CSR-backed expression (expressions.py:2033): it reads self.data and densifies, so chained group-sums always lose the backing.

Implementation ideas

The kernel already exists. CSRLinearExpression.contracted (added in #961) evaluates (I_kept (x) C^T) @ A in chunks. A sum over a dim is that contraction against a ones vector, so the work is wiring rather than new linear algebra:

• sum(dim) maps to contracted with an all-ones operand over the summed dims and no new output dims.
• sum() over everything maps to the existing full-contraction case, which already yields a zero-dim grid.
• The result is compact canonical form, so nterm may differ from the dense path. That is the same contract @ established, v1 only; see Zero culling forces model rebuilding #925.

Chained groupby needs CSRLinearExpression.from_grouper to accept an existing CSR source rather than a dense Dataset.