Add DGKS iterative reorthogonalization utility

Add internal::dgks_orthogonalize() implementing the Daniel-Gragg-Kaufman-Stewart criterion for Classical Gram-Schmidt with iterative refinement.

Uses BLAS-2 operations (GEMV) for orthogonalization and checks whether the norm dropped by more than 1/sqrt(2) to decide if reorthogonalization is needed. The primary benefit is numerical robustness — CGS with DGKS reorthogonalization achieves the same stability as MGS while using block operations that can be more efficient when paired with an optimized BLAS in the tall-matrix regime.

API

// Orthogonalize v against first numCols columns of Q.
// Returns stable norm of v after orthogonalization.
RealScalar internal::dgks_orthogonalize(Q, v, numCols, &coeffs, maxIter=2);
RealScalar internal::dgks_orthogonalize(Q, v, numCols, maxIter=2);

Benchmark

Orthogonalizing a random vector against an orthonormal basis (g++ -O3 -march=native -DNDEBUG, double precision):

Size (n x k)	MGS (us)	DGKS (us)	Speedup	Winner
8 x 4	0.1	0.2	0.3x	MGS
16 x 8	0.1	0.1	0.5x	MGS
32 x 16	0.2	0.2	0.9x	~tie
64 x 32	0.4	0.4	1.2x	~tie
128 x 64	1.5	1.6	0.9x	~tie
256 x 128	5.7	6.2	0.9x	MGS
512 x 256	21.4	27.1	0.8x	MGS
1024 x 512	96.6	168.1	0.6x	MGS

Tall matrices (many rows, moderate basis):

Size (n x k)	MGS (us)	DGKS (us)	Speedup	Winner
100 x 10	0.2	0.3	0.8x	MGS
500 x 10	0.9	1.0	0.9x	~tie
1000 x 10	2.4	2.5	0.9x	~tie
1000 x 50	8.2	6.5	1.3x	DGKS
1000 x 100	18.1	14.0	1.3x	DGKS

DGKS wins in the tall-matrix regime where GEMV arithmetic intensity is high. For square-ish cases, MGS loop has less overhead. The primary value of DGKS is numerical robustness — guaranteed orthogonality via iterative reorthogonalization, not raw speed.

Tests

13 subtests covering: basic orthogonality (float/double/complex), coefficient accumulation, vector-in-span, already-orthogonal, nearly-dependent, extreme magnitudes, single column, zero columns.

Reference

J.W. Daniel, W.B. Gragg, L. Kaufman, G.W. Stewart, Reorthogonalization and stable algorithms for updating the Gram-Schmidt QR factorization, Math. Comp. 30 (1976), 772-795.