LLM: fix baichuan7b quantize kv abnormal output. (intel-analytics#10504)

* fix abnormal output.

* fix style.

* fix style.

python/llm/src/bigdl/llm/transformers/models/baichuan.py

@@ -128,7 +128,8 @@ def baichuan_attention_forward_7b_quantized(
                 bsz, self.num_heads, kv_seq_len, self.head_dim,
                 device=device, new_layout=True
             )
-            key_states, value_states = append_kv_cache(k_cache, v_cache, key_states, value_states)
+            key_states, value_states = append_fp8_kv_cache(k_cache, v_cache, key_states,
+                                                           value_states, new_layout=True)
             past_key_value = (key_states, value_states)
     else:
         k_cache, v_cache = past_key_value
@@ -185,7 +186,7 @@ def baichuan_attention_forward_7b_quantized(
     if not output_attentions:
         attn_weights = None
 
-    return attn_output, attn_weights, past_key_value
+    return attn_output.to(hidden_states.dtype), attn_weights, past_key_value
 
 
 def baichuan_attention_forward_7b_origin(