ggml-cuda: enable the shared-expert overlap by default on RDNA3.5

GGML_CUDA_GRAPH_OPT still forces the optimization on ("1") or off ("0"), but when
it is unset the default is now on for RDNA3.5 - the architecture where the
decode-time shared-expert overlap has been tuned - and off everywhere else.

Co-Authored-By: Claude Opus 4 (1M context) <noreply@anthropic.com>

Author: roberteg16

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -4525,11 +4525,17 @@ static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph
     GGML_UNUSED(cgraph);
 #endif
 
-    static bool enable_graph_optimization = [] {
-        const char * env     = getenv("GGML_CUDA_GRAPH_OPT");
-        return env != nullptr && atoi(env) == 1;
+    // GGML_CUDA_GRAPH_OPT: "1"/"0" force the optimization on/off; when unset it defaults on for
+    // RDNA3.5, the only architecture where the decode-time overlap has been tuned, and off elsewhere.
+    static const int graph_opt_env = [] {
+        const char * env = getenv("GGML_CUDA_GRAPH_OPT");
+        return env == nullptr ? -1 : atoi(env);
     }();
 
+    const bool enable_graph_optimization = graph_opt_env >= 0 ?
+        graph_opt_env == 1 :
+        GGML_CUDA_CC_IS_RDNA3_5(ggml_cuda_info().devices[cuda_ctx->device].cc);
+
     if (!enable_graph_optimization) {
         return;
     }