fix a typo in multihead-attention.md

In `MultiHeadAttention.forward()`, the sizes of the 3rd dim of tensors `queries`, `keys` and `values` should be `query_size`, `key_size` and `value_size` respectively, instead of `num_hiddens` as written in the comments. This commit fixed this typo.

chapter_attention-mechanisms/multihead-attention.md

@@ -109,7 +109,7 @@ class MultiHeadAttention(nn.Block):
 
     def forward(self, queries, keys, values, valid_lens):
         # queries，keys，values的形状:
-        # (batch_size，查询或者“键－值”对的个数，num_hiddens)
+        # (batch_size，查询或者“键－值”对的个数，query_size或key_size或value_size)
         # valid_lens　的形状:
         # (batch_size，)或(batch_size，查询的个数)
         # 经过变换后，输出的queries，keys，values　的形状:
@@ -437,4 +437,4 @@ attention(X, Y, Y, valid_lens, training=False).shape
 
 :begin_tab:`paddle`
 [Discussions](https://discuss.d2l.ai/t/11843)
-:end_tab:
+:end_tab: