[FEAT] Support transpose flags transa, transb != 'N' in GEMM #13

Summary

qblas_gemm and the higher-level wrappers currently reject any value other than 'N' for the transpose flags, which makes it impossible for
callers to handle mixed-layout operands (e.g. row-major × col-major) without doing a full copy into a contiguous buffer first.

Reference (in the numpy-quaddtype tree at src/csrc/quadblas_interface.cpp, but the upstream QBLAS-side QuadBLAS::gemm shares the same limitation):

// For now, we only support no transpose
// TODO: Implement transpose support if needed
if ((transa != 'N' && transa != 'n') || (transb != 'N' && transb != 'n')) {
    return -1;  // Transpose not implemented yet
}

QuadBLAS::gemm(qblas_layout, m, n, k, *alpha, A, lda, B, ldb, *beta, C, ldc);

qblas_gemv already accepts 'T' (implemented as a layout swap), but
the same T / C semantics aren't available for gemm.

Proposed API

Match the BLAS convention so the wrapper can pass transa/transb
through directly:

// Each of 'N', 'T', 'C' supported (C == T for real-valued types).
QuadBLAS::gemm(Layout layout,
               Trans transa, Trans transb,
               size_t m, size_t n, size_t k,
               Sleef_quad alpha,
               const Sleef_quad *A, size_t lda,
               const Sleef_quad *B, size_t ldb,
               Sleef_quad beta,
               Sleef_quad *C, size_t ldc);

Layout already exists; adding a Trans enum (or a char) and honoring it in the kernel would let the C wrapper drop the "transpose not implemented yet" guard.

Notes / acceptance

• Implementation can reuse the same trick qblas_gemv uses (swap dims + flip layout) when that's enough, or do a real transposed kernel where it matters for performance.
• Existing tests in tests/ should be extended with transa/transb variants of the gemm cases.