accuracy gap of stft in float16

### 🐛 Describe the bug

pytest -v test/inductor/test_torchinductor_opinfo.py -k test_comprehensive_stft_xpu_float16
```
inductor/test_torchinductor_opinfo.py::TestInductorOpInfoXPU::test_comprehensive_stft_xpu_float16 - Exception: Tensor-likes are not close!
2026-04-08T06:00:29.3521635Z
2026-04-08T06:00:29.3521868Z Mismatched elements: 11 / 100 (11.0%)
2026-04-08T06:00:29.3522161Z Greatest absolute difference: 0.000244140625 at index (26,) (up to 1e-05 allowed)
2026-04-08T06:00:29.3522524Z Greatest relative difference: 0.00458526611328125 at index (52,) (up to 0.001 allowed)
2026-04-08T06:00:29.3522578Z
2026-04-08T06:00:29.3523228Z Caused by sample input at index 0: SampleInput(input=Tensor[size=(100,), device="xpu:0", dtype=torch.float16], args=(), kwargs={'n_fft': '10', 'return_complex': 'True'}, broadcasts_input=False, name='')
2026-04-08T06:00:29.3523234Z
2026-04-08T06:00:29.3523543Z To execute this test, run the following from the base repo dir:
2026-04-08T06:00:29.3524085Z PYTORCH_OPINFO_SAMPLE_INPUT_INDEX=0 python test/inductor/test_torchinductor_opinfo.py TestInductorOpInfoXPU.test_comprehensive_stft_xpu_float16

```

### Versions

https://github.com/pytorch/pytorch/pull/179550