修复 benchmark CI NPU test 模型问题

[xuedinge233]

[hipudding]

• 验证所有失败的模型的失败原因，并评估是否可以修复

[xuedinge233]

失败原因总结：

• npu out of memory
• No module named 'triton'
• not_implemented
• corrupted download, the hash of https://doctr-static.mindee.com/models?id=v0.3.1/crnn_vgg16_bn-9762b0b0.pt&src=0 does not match its expected value"
• ("<class '_pickle.UnpicklingError'>", "('pickle data was truncated',)")

[noemotiovon]

• 当前所有模型通过率为80%，目标是90%