BlockSparseMatrix: add block-sparse data structure and operations

Adds BlockSparseMatrix<Scalar, Options, BR, BC>, a sparse matrix whose nonzero entries are fixed-size dense blocks. It supports triangular and selfadjoint views with the same API as SparseMatrix, and plugs into the Eigen product evaluator so expressions like y.noalias() = A * x and A.triangularView<Upper>().solveInPlace(x) work without extra glue.

---

Benchmarks (nB=200 blocks, sparsity=5%)

Speedup is Sm time / BSM time; values >1 mean BSM is faster. For TriMV_DiagT the baseline is Sm_TriMV; for SymmMV_DiagNSA/SymmMV_DiagSA it is Sm_SymmMV.

Sparsity level (1%, 5%, 10%) has little effect on the relative speedups: both Sm and BSM traverse the same block sparsity pattern, so the gains come from within-block compute density rather than from how many blocks there are. Results at 5% are representative.

B=2

Operation	float	double	cf	cd
SpMV	1.46x	1.87x	2.50x	2.21x
TriMV	1.32x	1.51x	1.49x	1.50x
TriMV_DiagT	1.86x	2.36x	2.74x	2.05x
SymmMV_DiagNSA	0.87x	0.85x	1.54x	1.36x
SymmMV_DiagSA	1.03x	1.36x	1.68x	1.70x
TriSolve	2.60x	3.08x	2.01x	2.71x
SpSp_Add	10.38x	11.99x	9.90x	10.33x
SpSp_Mul	1.59x	1.84x	1.79x	1.99x

B=3

Operation	float	double	cf	cd
SpMV	1.87x	2.36x	1.65x	1.48x
TriMV	1.97x	2.01x	1.39x	1.35x
TriMV_DiagT	2.64x	2.63x	1.83x	1.51x
SymmMV_DiagNSA	0.97x	1.04x	1.30x	1.40x
SymmMV_DiagSA	1.30x	1.53x	1.42x	1.38x
TriSolve	2.95x	3.22x	1.79x	2.34x
SpSp_Add	21.18x	14.45x	14.31x	10.42x
SpSp_Mul	3.28x	3.85x	3.51x	3.14x

B=4

Operation	float	double	cf	cd
SpMV	4.21x	3.69x	5.54x	2.60x
TriMV	4.31x	4.09x	3.45x	2.54x
TriMV_DiagT	6.03x	5.42x	6.32x	3.17x
SymmMV_DiagNSA	2.19x	2.13x	3.15x	2.25x
SymmMV_DiagSA	2.22x	2.17x	3.80x	2.51x
TriSolve	8.14x	7.57x	2.88x	3.03x
SpSp_Add	35.61x	18.75x	19.28x	9.92x
SpSp_Mul	7.06x	7.48x	7.55x	6.52x