Avoid all-zeor attnetion mask used in testing (#26469)

fix

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

tests/test_modeling_common.py

@@ -2960,7 +2960,8 @@ def ids_tensor(shape, vocab_size, rng=None, name=None):
 def random_attention_mask(shape, rng=None, name=None):
     attn_mask = ids_tensor(shape, vocab_size=2, rng=None, name=None)
     # make sure that at least one token is attended to for each batch
-    attn_mask[:, -1] = 1
+    # we choose the 1st token so this property of `at least one being non-zero` still holds after applying causal mask
+    attn_mask[:, 0] = 1
     return attn_mask