Let reward model take text inputs instead of message lists (#1907)

Co-authored-by: Kyle Corbitt <kyle@corbt.com>

python/sglang/srt/models/llama_reward.py

@@ -52,7 +52,12 @@ class LlamaForSequenceClassification(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
+        get_embedding: bool = True,
     ) -> EmbeddingPoolerOutput:
+        assert (
+            get_embedding
+        ), "LlamaForSequenceClassification is only used for embedding"
+
         hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
         scores = self.score(hidden_states)