Make the sparse path observable and controllable (.is_sparse, densify warning, per-call sparse=)

Start by reading the sparse paths at linopy/expressions.py:2570, 3291, 3319, 2606-2609 and linopy/constraints.py:1427, then inspect Grid and the existing sparse_groupby option. Define how repr(), shape, sizes, coords, dims and isnull avoid densifying before implementing the public state, warnings and per-call controls. Done means sparse fallbacks are observable and controllable without changing default behavior.

[!NOTE]
The following content was generated by AI.

Describe the feature you'd like to see

Make the sparse path observable and controllable. Child of #756.

Right now a user cannot tell whether their build stayed sparse. There is no .is_sparse property, no marker in the repr, and no warning when the backing is dropped. Four separate code paths fall back to dense by returning None, all silently: _sparse_matmul (linopy/expressions.py:2570), _try_csr_merge (3319), _aligned (3291) and csr_rhs (linopy/constraints.py:1427).

Combined with the fact that reading repr() or .sizes destroys the backing in place (filed separately as a bug), this makes the feature very hard to use deliberately on anything larger than a toy model.

Three parts:

1. Expose the state. A public .is_sparse on LinearExpression, and a marker in __repr__ — provided the repr itself stops densifying.
2. Explain the fallbacks. An opt-in warning, for example options["warn_on_densify"], naming the operation that dropped the backing. Off by default, so nothing changes for existing users.
3. Per-call control. A sparse= flag on @/dot and merge, mirroring groupby(...).sum(sparse=True). Today @ contracts sparsely but returns a dense object unless the input was already CSR-backed or the global options["sparse_groupby"] is set (expressions.py:2606-2609), which is a surprising coupling between an unrelated option and the result's representation. #961 explicitly deferred this flag.

Implementation ideas

Serving shape, sizes, coords, dims and isnull from the existing Grid is a prerequisite for a non-densifying repr, and is worth doing on its own: Grid already holds all of it.