[GPT2] Fix encoder_attention_mask being silently discarded in cross-attention (#47946)

* [Bug] Fix encoder_attention_mask being silently discarded in GPT2 cross-attention

The new attn mask interface (#42848) rewired GPT2Model and its decision_transformer
copy to build the cross-attention mask via create_bidirectional_mask, but added an
encoder_attention_mask = None line directly above the call that overwrote the incoming
mask before it was passed in. As a result, encoder padding was ignored in cross-attention.
Remove the line and add a regression test.

* Update tests/models/gpt2/test_modeling_gpt2.py

---------

Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>