Also use column-major B matrix in the example 00_bmg_gemm.cpp

[sanchitintel]

Background

For a GEMM, [K, N] is the canonical shape of the canonical B matrix, so RowMajor tends to mean [K, N].

In the current GEMM kernel & MMA collective examples such as those based on 00_bmg_gemm.cpp, the canonical B matrix is always shaped [N, K] in the code (maybe to conform to the CuTe convention that B is always shaped [N, K]?). it's row-major, but discontiguous (physical layout in the memory is [K, N], and stride is [N, 1], so it's actually transposed).

We could also support the column-major case for B, in which case it's really shaped [N, K].

Summary of changes

Also add a case for when B (shaped [N, K]) is contiguous (with a stride of [K, 1]), because such use-cases are common in LLMs. To support it, the code-changes needed are:

1. B layout is cutlass::layout::ColumnMajor, and
2. the copy-atom used is different.

00_bmg_gemm.cpp now runs both these cases.

Thanks!