Extract transpose-flip helper in bgemm_blas.rs

[shinaoka]

Summary

strided-einsum2/src/bgemm_blas.rs lines 408-437: when C is row-major, the GEMM is rewritten as C^T = B^T * A^T with inline transpose flag flipping logic. This is correct but adds cognitive load as an inline closure within do_batch.

Proposal

Extract the transpose-flip logic into a standalone helper function:

fn flip_transpose(t: CBLAS_TRANSPOSE) -> CBLAS_TRANSPOSE {
    match t {
        CblasNoTrans => CblasTrans,
        CblasTrans => CblasNoTrans,
        other => other,
    }
}

And separate the col-major vs row-major C dispatch into clearly named branches.

Risk

Low. Purely cosmetic refactoring for readability.