Method build_rope_2d

subprojects/llama.cpp/tools/mtmd/clip.cpp:646–711

implementation of the 2D RoPE without adding a new op in ggml this is not efficient (use double the memory), but works on all backends TODO: there was a more efficient which relies on ggml_view and ggml_rope_ext_inplace, but the rope inplace does not work well with non-contiguous tensors ; we should fix that and revert back to the original implementation in https://github.com/ggml-org/llama.cpp/pu

// this is not efficient (use double the memory), but works on all backends
// TODO: there was a more efficient which relies on ggml_view and ggml_rope_ext_inplace, but the rope inplace does not work well with non-contiguous tensors ; we should fix that and revert back to the original implementation in https://github.com/ggml-org/llama.cpp/pull/13065
ggml_tensor * clip_graph::build_rope_2d(
    ggml_context * ctx0,
    ggml_tensor * cur,
    ggml_tensor * pos_a, // first half
    ggml_tensor * pos_b, // second half
    const float freq_base,
    const bool interleave_freq
) {
    const int64_t n_dim  = cur->ne[0];
    const int64_t n_head = cur->ne[1];
    const int64_t n_pos  = cur->ne[2];

    // for example, if we have cur tensor of shape (n_dim=8, n_head, n_pos)
    // we will have a list of 4 inv_freq: 1e-0, 1e-1, 1e-2, 1e-3
    // first half of cur will use 1e-0, 1e-2 (even)
    // second half of cur will use 1e-1, 1e-3 (odd)
    // the trick here is to rotate just half of n_dim, so inv_freq will automatically be even
    //  ^ don't ask me why, it's math! -2(2i) / n_dim == -2i / (n_dim/2)
    // then for the second half, we use freq_scale to shift the inv_freq
    //  ^ why? replace (2i) with (2i+1) in the above equation
    const float freq_scale_odd = interleave_freq
                                ? std::pow(freq_base, (float)-2/n_dim)
                                : 1.0;

    // first half
    ggml_tensor * first;
    {
        first = ggml_view_3d(ctx0, cur,
            n_dim/2, n_head, n_pos,
            ggml_row_size(cur->type, n_dim),
            ggml_row_size(cur->type, n_dim*n_head),
            0);
        first = ggml_rope_ext(
            ctx0,
            first,
            pos_a,      // positions
            nullptr,    // freq factors
            n_dim/2,    // n_dims
            0, 0, freq_base,
            1.0f, 0.0f, 1.0f, 0.0f, 0.0f
        );
    }

    // second half
    ggml_tensor * second;
    {
        second = ggml_view_3d(ctx0, cur,
            n_dim/2, n_head, n_pos,
            ggml_row_size(cur->type, n_dim),
            ggml_row_size(cur->type, n_dim*n_head),
            n_dim/2 * ggml_element_size(cur));
        second = ggml_rope_ext(
            ctx0,
            second,
            pos_b,      // positions
            nullptr,    // freq factors
            n_dim/2,    // n_dims
            0, 0, freq_base,