[bug fix] update quantizable add ops detection on IPEX backend

[xin3he]

Type of Change

bug fix

Description

IPEX backend got wrong number of quantizable add ops when detecting.
Quantizable add ops should be detected in gptj but not in opt model.

Expected Behavior & Potential Risk

add ops can only be detected if both input inf_dtypes are not float32.

        "q_op_infos": {
            "0": {
                "op_type": "<method 'add' of 'torch._C._TensorBase' objects>",
                "op_type_is_module": false,
                "fqn": "",
                "input_tensor_infos": [
                    {
                        "id": 0,
                        "orig_dtype": "torch.float32",
                        "inf_dtype": "torch.float32",
                        "force_dtype": "torch.float32",
                        "scale": [
                            0.007949741557240486
                        ],
                        "zero_point": [
                            219
                        ]
                    },
                    {
                        "id": 0,
                        "orig_dtype": "torch.float32",
                        "inf_dtype": "torch.float32",
                        "force_dtype": "torch.float32",
                        "scale": [
                            0.007949741557240486
                        ],
                        "zero_point": [
                            219
                        ]
                    }
                ],

How has this PR been tested?

local test on models

Dependency Change?

N/A