context : reserve new scheduler when graph topology changes (#18547)

* context : reserve new scheduler when graph topology changes

* cont : fix

* cont : fix reserve

* cont : reserve only when changes occur + timing

* context : add comments

* llama : reserve on sampler changes

* common : allow null common_sampler

* server : task declares needs (embd, logits, sampling)

* server : do not init sampler if not needed

* llama : fix need_reserve when unsetting a sampler

* server : consolidate slot reset/clear logic

src/llama-context.h

@@ -40,6 +40,14 @@ struct llama_context {
 
     ~llama_context();
 
+    // reserve a new backend scheduler (if needed)
+    // for example, when:
+    //   - changing loras
+    //   - changing samplers
+    //   - changing attention type
+    //   - etc.
+    void sched_reserve();
+
     void synchronize();
 
     const llama_model   & get_model()   const;
@@ -314,6 +322,8 @@ private:
 
     ggml_backend_sched_ptr sched;
 
+    bool sched_need_reserve = true;
+
     ggml_backend_t backend_cpu = nullptr;
     std::vector<ggml_backend_ptr> backends;