model : add LLADA 2.0 diffusion support (Diffusion masking/kv cache)

[wsbagnsv1]

Since llada2 uses kv cache outside its diffusion blocks in the transformers implementation I implemented that in my pr with a diffusion cli arg toggle, but got advised that its not recommended to add optimizations in the pr for support only later on in another optimization pr, so now I have an issue. I dont wanna use a weird cli arg just for llada2, which would be needed to add proper masking support with kv cache handling in diffusion cli itself which is also not good (without the full so i set out and tried to implement it inside llada2.cpp, but there I got issues with kv cache allocation crashes so I thought of another way to implement the proper masking, and found that i can do it in llama-kv-cache.cpp

undervoid llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn, int32_t block_length) const

since other models might need the same logic in the future and it would be gated by block size so only diffusion models could access that code:

if (block_length > 0) {
const int32_t block_p0 = p0 / block_length;
const int32_t block_p1 = p1 / block_length;

                    // mask if key is in a future block
                    if (block_p0 > block_p1) {
                        continue;
                    }
                }

Now my question is basically if that is a good addition for future models or if it might cause issues with future block based diffusion models that use kv cache like llada2 which might rely on something else? That way i can use kv cache later without issues (;
Also should i open another pr for that extra support?