DeepseekV4: fix rollback with multi-seq (#26756)

* DeepseekV4: fix rollback with multi-seq

* fix model loading

* make pending rollback single use

* only clear cache for seq_id for full load

* add assert for compress ratio

* make graph topology static

* pass true instead of flags in clear_compressed

* cont : clean-up + TODOs

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

src/models/deepseek4.cpp

@@ -1,3 +1,4 @@
+#include "llama-hparams.h"
 #include "models.h"
 
 #include "llama-kv-cache-dsv4.h"
@@ -58,6 +59,7 @@ void llama_model_deepseek4::load_arch_hparams(llama_model_loader & ml) {
     if (n_compress_ratios < hparams.n_layer_all) {
         throw std::runtime_error("DeepSeek-V4 compress_ratios is shorter than block_count");
     }
+    GGML_ASSERT(n_compress_ratios <= LLAMA_MAX_LAYERS);
     ml.get_arr(LLM_KV_ATTENTION_COMPRESS_RATIOS, hparams.dsv4_compress_ratios);
 
     ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func);