Fix Core Attn in ernie 4.5

[cheng221]

Fix Error in These two situation (This PR only affects the self_attention in modeling when use_flash_attention is set to False):

1. When using the ./tools/inference/infer.py script for inference, setting use_flash_attention results in the following error

2. This also exists in ./examples/post-training/sft/scripts/run_0.3b_sft_8k.sh