BlockSparseMatrix might benefit from caching column-block partition

[sandwichmaker]

There is a design issue as to who should cache this partitioning, but one place where this could be done for sure is inside CgnrLinearOperator, since within its lifetime, the block sparsity structure is guaranteed to remain constant. This should have a nice effect on the performance of CGNR when calling LeftMultiplyAndAccumulate.

Other methods that may benefit from this are:
SquareColumnNorm
ScaleColumns

But the frequency with which they are called is much smaller.

[sandwichmaker]

cc: @DmitriyKorchemkin