Accuracy issue with 64 bit indexing depthwise_conv

### 🐛 Describe the bug

Cases:
op_ut,third_party.torch-xpu-ops.test.xpu.nn.test_convolution_xpu.TestConvolutionNNDeviceTypeXPU,test_depthwise_conv_64bit_indexing_xpu

Error:
```Python
E AssertionError: Tensor-likes are not close! Mismatched elements: 4311067 / 2151680000 (0.2%) Greatest absolute difference: 0.00390625 at index (0, 0, 5, 18335) (up to 0.001 allowed)
E Greatest relative difference: inf at index (0, 0, 6063, 19660) (up to 0.0001 allowed)
E To execute this test, run the following from the base repo dir: PYTORCH_TEST_WITH_SLOW=1 python ../../test/nn/test_convolution.py TestConvolutionNNDeviceTypeXPU.test_depthwise_conv_64bit_indexing_xpu
E This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
```

### Versions

PyTorch2.11.0.dev20260203+xpu