gpu: fix (avoid) jit matmul post-ops with layout not matching dst

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: karturov

src/gpu/intel/gemm/jit/gen_kernel.cpp

@@ -401,9 +401,28 @@ status_t gen_desc_t::transfer_post_ops(
 
             bool trans = is_multi_row && !src_rmd.inner_dim.is_innermost();
 
+            // Reject col-only binary post-ops with non-unit stride in the
+            // col direction. Non-unit stride requires Scattered access with
+            // correct multi-block offsets, which is not yet supported.
+            {
+                int rmd_ndims = src_rmd.ndims();
+                bool col_only = is_multi_col && !is_multi_row;
+                if (col_only
+                        && !src_rmd.is_inner_dim(rmd_ndims - 2, rmd_ndims))
+                    return status::unimplemented;
+            }
+
             if (swap_ab) {
                 trans = !trans;
                 std::swap(is_multi_row, is_multi_col);
+
+                // After swap, original row-only becomes col-only. Col
+                // direction is now ndims-1 instead of ndims-2.
+                int rmd_ndims = src_rmd.ndims();
+                bool col_only = is_multi_col && !is_multi_row;
+                if (col_only
+                        && !src_rmd.is_inner_dim(rmd_ndims - 1, rmd_ndims))
+                    return status::unimplemented;
             }
 
             problem_.Tbinary.push_back(T);