perf: disable llama perf counters in hot path

crates/inference-core/src/engine.rs

@@ -232,7 +232,7 @@ impl InferenceEngine {
             LLAMA_FLASH_ATTN_TYPE_DISABLED
         };
         cparams.offload_kqv = true;
-        cparams.no_perf = false;
+        cparams.no_perf = true;
 
         // 6. Init context
         let ctx_ptr = unsafe { (lib.init_from_model)(model_ptr, cparams) };