[FIX] autoround depend causing torch-cpu to be installed

[Qubitium]

I don't know why our ci did not pick up on this bug there is some strange dependency override issue with intel extension for transformers is overriding the torch[cuda]=2.4.1 installed with a torch[cpu] 2.3.0.

EDIT: source of bug is actually autoround 0.3 setup code

[Qubitium]

@LRL-ModelCloud we need to fix this. First, check if auto-round depends on intel-extension-for-transformers and test pip install -vv . on a clean python 3.12 conda env.

We may need to move auto-round to optional/secondary package if it becomes an issue.

First check who is using intel-extension-for-transformers?

There is also a second issue where the vllm/sglang depends are too old. Update the versions and flashinfer (do we need to add flashinfer since they are auto-depen by vllm and sglang latest?).