[ENH] vectorize CLA `_reduce_matrix` using `numpy` advanced indexing (#693)

Replace nested loop implementation with np.ix_ based vectorized
selection. Performance improvement: 10-36x speedup depending on matrix
size.

Benchmarks (100 iterations):
- n=20:  10x faster
- n=50:  18x faster
- n=100: 12x faster
- n=200: 34x faster
- n=500: 37x faster

All existing CLA tests pass with identical numerical results.