Only emit qk_norm_epsilon GQA attribute when q/k norm weights are present

The unconditional qk_norm_epsilon attribute broke plain GQA models (e.g.
phi-2): the current onnxruntime nightly does not recognize the attribute,
so model load failed with "Unrecognized attribute: qk_norm_epsilon for
operator GroupQueryAttention". It also failed the
test_plain_gqa_omits_qk_norm_epsilon_attribute unit test. Restore the
guard so the attribute is only added on the fused Q/K norm path.

Author: tianleiwu

src/python/py/models/builders/base.py

@@ -2751,8 +2751,9 @@ def make_group_query_attention(self, name, **kwargs):
             "softcap": self.attention_attrs["softcap"],
             "do_rotary": self.attention_attrs["use_rope_in_attn"],
             "rotary_interleaved": self.rope_attrs["interleaved"],
-            "qk_norm_epsilon": kwargs.get("qk_norm_epsilon", self.attention_attrs["qk_norm_epsilon"]),
         }
+        if q_norm_weight:
+            attributes["qk_norm_epsilon"] = kwargs.get("qk_norm_epsilon", self.attention_attrs["qk_norm_epsilon"])
         self.make_node(
             "GroupQueryAttention",
             inputs=inputs,