[raashish1601]

Fixes #669.

Grid.to_LRCGeometry() converts the point cloud of each grid dimension with geom.to_LRCGeometry(rank=0, ...) and then reads cost_1 / cost_2. For a squared Euclidean (or negative dot product) point cloud, PointCloud.to_LRCGeometry() returns the point cloud itself when the factors would not be smaller than the cost matrix (n * m <= (n + m) * d). With 2 points on a dimension that is always the case (4 <= 4), so any grid with a dimension of size 2 fails with AttributeError: 'PointCloud' object has no attribute 'cost_1'. SinkhornOutput.primal_cost hits this through transport_cost_at_geom(), since Grid.can_LRC is True.

When the slice conversion does not return an LRCGeometry, the grid now factorizes that slice with the exact SVD from Geometry.to_LRCGeometry(rank=0), which is what the comment above the loop already asks for.

Tests:

• test_conversion_grid in tests/geometry/lr_cost_test.py is parametrized with grids that have size-2 dimensions ([2], [2, 3], [3, 2, 2]) in addition to the existing [6, 7, 11].
• test_primal_cost_grid_with_two_points in tests/solvers/linear/sinkhorn_grid_test.py compares primal_cost on grids of size (2,), (2, 2) and (2, 3) with the same problem on a PointCloud.

Both fail on main and pass with the change. tests/solvers/linear/sinkhorn_grid_test.py and the grid conversion tests pass locally (CPU), and ruff, isort and yapf at the pre-commit versions report no changes.