CUDA: fix MMQ for non-contiguous src0, add tests

[JohannesGaessler]

Fixes #10011 .

The problem is that on master the MMQ code calculates the stride for src0 based on src0->nb[1] but in ggml_cuda_op_mul_mat any non-contiguous matrices are made contiguous first. This case was not covered by the tests, I extended test-backend-ops with new test cases for non-contiguous inputs.

The temporary buffer for src0 also did not get its padding cleared. This did not seem to result in incorrect results but it would in principle be possible so I also fixed that.

Long-term I think we should refactor the code in such a way that we don't need ggml_cuda_op_mul_mat and instead handle batched matrix multiplication in the kernels themselves.

[slaren]

I find this a bit hard to follow. Would this be equivalent to what you are doing here?

        // C^T = A * B^T: (k, m) * (k, n) => (m, n)
        ggml_tensor * a;
        ggml_tensor * b;

        // If the permutation is not {0, 1, 2, 3}, replace a and b with views that have the same data in a different order.
        // This test only works correctly if exactly 2 indices != 0 are swapped.
        if (per[0] != 0 || per[1] != 1 || per[2] != 2 || per[3] != 3) {
            // create a tensor with the permuted dimensions, them permute it back to the dimensions given by m,n,k
            int64_t ne_a[4] = {k, m, bs[0], bs[1]};
            int64_t ne_b[4] = {k, n, bs[0]*nr[0], bs[1]*nr[1]};
            a = ggml_new_tensor_4d(ctx, type_a, ne_a[per[0]], ne_a[per[1]], ne_a[per[2]], ne_a[per[3]]);
            b = ggml_new_tensor_4d(ctx, type_b, ne_b[per[0]], ne_b[per[1]], ne_b[per[2]], ne_b[per[3]]);
            ggml_set_name(a, "a");
            ggml_set_name(b, "b");
            a = ggml_permute(ctx, a, per[0], per[1], per[2], per[3]);
            b = ggml_permute(ctx, b, per[0], per[1], per[2], per[3]);
            ggml_set_name(a, "a_permuted");
            ggml_set_name(b, "b_permuted");
        } else {
            a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0]      , bs[1]);
            b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
            ggml_set_name(a, "a");
            ggml_set_name(b, "b");
        }
        ggml_set_param(ctx, a);
        ggml_set_param(ctx, b);

[JohannesGaessler]

Thanks, that's probably a better way to do it.