feat(local-inference): add TTL support for automatic model unloading

Add globalTTL and per-model ttl options to llama-swap config,
allowing idle models to be automatically unloaded from memory.

machines/zix790prors/configuration.nix

@@ -30,6 +30,7 @@ with lib;
       enable = true;
       host = "zix790prors.oglehome";
       openFirewall = true;
+      globalTTL = 900;
       models = {
         "Qwen3.6-35B-A3B" = {
           hf-model = "unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q8_K_XL";