support mode parameter in PyTorch 2.1

[haoyang9804]

修复了--compile_with_dynamo flag下oneflow无法支持PyTorch 2.1的bug
bug复现如下

python ~/onediff2/examples/experimental/text_to_image_sdxl_torch_compile.py --compile=false --compile_with_dynamo --base=/share_nfs/hf_models/stable-diffusion-xl-base-1.0

bug触发位置

if cmd_args.compile_with_dynamo:
    print("unet is compiled to oneflow with torch.compile.")
    from onediff.infer_compiler import oneflow_backend

    base.unet = torch.compile(
        base.unet, fullgraph=True, mode="reduce-overhead", backend=oneflow_backend
    )