Refactor CUDA step buffers to remove loop-time allocations

AMSS_NCKU_source/bssn_rhs_cuda.cu

@@ -3626,8 +3626,8 @@ int bssn_cuda_rk4_substep(void *block_tag,
     } else {
         download_state_outputs(state_host_out, all);
     }
-    if (RK4 == 3 && !use_resident_state) {
-        release_step_ctx(block_tag);
+    if (RK4 == 3) {
+        ctx.matter_ready = false;  /* invalidate matter cache for next timestep */
     }
     if (profile) {
         cuda_profile_sync();