Add mixed 8-bit attn + 4-bit MLP variant (92.2%)

mlx/llm-mixed8_4/generation_config.json

@@ -0,0 +1,9 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": [
+    151643,
+    151645
+  ],
+  "pad_token_id": 151643,
+  "do_sample": false
+}