Update config.json

Set key value heads.
Previously was explicitly null, as the logic is per block. In vLLM code:
```
num_kv_heads = (config.num_attention_heads //
block_config.attention.n_heads_in_group)
```

Setting default number here: 64/8 = 8

config.json

@@ -1461,7 +1461,7 @@
   "model_type": "nemotron-nas",
   "num_attention_heads": 64,
   "num_hidden_layers": 80,
-  "num_key_value_heads": null,
+  "num_key_value_heads": 8,
   "pretraining_tp": 1,
   "quantization_config": {
     "config_groups": {