[Release/2.5.0] Torchbench basic_gnn models performance regression

[mengfei25]

🐛 Describe the bug

Category	Name	Speedup [2.5.0]	Speedup [main]	2.5.0 vs. main [Eager]	2.5.0 vs. main [Inductor]
torchbench_bfloat16_training	basic_gnn_gcn	0.663553	1.200854	0.700546215	0.387099133
torchbench_amp_bf16_training	basic_gnn_gin	0.672736	1.136112	0.673174765	0.398612899
torchbench_bfloat16_training	basic_gnn_gin	0.659017	1.162534	0.706012305	0.400224089
torchbench_amp_bf16_training	basic_gnn_gcn	0.648738	1.105941	0.687184282	0.403097956
torchbench_amp_bf16_training	basic_gnn_sage	0.661256	1.102461	0.695784714	0.417331604
torchbench_bfloat16_training	basic_gnn_edgecnn	0.912495	1.629383	0.756308349	0.423551484
torchbench_amp_bf16_training	basic_gnn_edgecnn	0.93919	1.634551	0.752725408	0.432505425
torchbench_bfloat16_training	basic_gnn_sage	0.684153	1.089889	0.724069202	0.454517953

Versions

pytorch: 2.5.0-rc1 (https://download.pytorch.org/whl/test/xpu)
torch-xpu-ops: 1206590 (main)