[RFC]: add `blas/ext/base/gfill-columns`

TODO: sthg like `gfillColumns.ndarray( M, N, alpha, strideA, offsetA, x, strideX1, strideX2, offsetX )` where `alpha` is a strided array of fill values. This would allow filling individual columns separately. Why not a loop over `gfill`? Because, for row-major, instead of filling columns one-by-one, one iterates across the rows and repeatedly cycles over `alpha` for better cache usage.

Start by reviewing the proposed gfillColumns.ndarray signature and comparing it with the existing gfill implementation and BLAS extension conventions. Determine the appropriate entry point and tests for a strided, per-column fill, then define completion as working behavior for the stated row-major cache-friendly case and coverage of the relevant layouts and offsets.

索引モデルが issue の本文から書いたものです。