[upstream_ut] test_torch.py nondeterministic_alert failures: XPU backward kernels raise stale non-deterministic error

## Description
25 `test_nondeterministic_alert_*` cases in `test/test_torch.py` fail on XPU. The tests do NOT expect a non-determinism error in the default (non-deterministic) mode, but the XPU backward kernels still raise:

RuntimeError: _backward_xpu does not have a deterministic implementation, but you set 'torch.use_deterministic_algorithms(True)'...

producing: "did not expect non-deterministic error message, but got one anyway".

This resembles previously fixed issues (#3117 stale alertNotDeterministic on scatter kernels, #2512/#2516/#2517) and looks like a stale/over-broad `alertNotDeterministic` on additional ops.

## Affected tests
```
test_nondeterministic_alert_AdaptiveAvgPool2d_xpu
test_nondeterministic_alert_AdaptiveAvgPool3d_xpu
test_nondeterministic_alert_AdaptiveMaxPool2d_xpu
test_nondeterministic_alert_AvgPool3d_xpu
test_nondeterministic_alert_MaxPool3d_xpu
test_nondeterministic_alert_FractionalMaxPool2d_xpu
test_nondeterministic_alert_FractionalMaxPool3d_xpu
test_nondeterministic_alert_ReflectionPad1d_xpu
test_nondeterministic_alert_ReflectionPad3d_xpu
test_nondeterministic_alert_ReplicationPad1d_xpu
test_nondeterministic_alert_ReplicationPad2d_xpu
test_nondeterministic_alert_ReplicationPad3d_xpu
test_nondeterministic_alert_interpolate_bilinear_xpu
test_nondeterministic_alert_interpolate_bicubic_xpu
test_nondeterministic_alert_interpolate_linear_xpu
test_nondeterministic_alert_interpolate_trilinear_xpu
test_nondeterministic_alert_grid_sample_2d_xpu
test_nondeterministic_alert_grid_sample_3d_xpu
test_nondeterministic_alert_CTCLoss_xpu
test_nondeterministic_alert_NLLLoss_xpu
test_nondeterministic_alert_EmbeddingBag_max_xpu
test_nondeterministic_alert_bincount_xpu
test_nondeterministic_alert_histc_xpu_float32
test_nondeterministic_alert_median_xpu_float64
test_nondeterministic_alert_put_accumulate_xpu
```

## Error log (example)
```
AssertionError: did not expect non-deterministic error message, but got one anyway: "adaptive_max_pool2d_backward_xpu does not have a deterministic implementation, but you set 'torch.use_deterministic_algorithms(True)'..."
```

## Reproduce
Using pytorch/pytorch#196421 on a PVC machine:
```
python test/test_torch.py TestTorchDeviceTypeXPU.test_nondeterministic_alert_AdaptiveMaxPool2d_xpu
```

## Expected
Match CUDA behavior: these ops should not raise a non-determinism error in the default mode (they should either be deterministic or not flagged when use_deterministic_algorithms is False).

_Filed with AI assistance; reproduced per the details above._

Run the listed TestTorchDeviceTypeXPU case from test/test_torch.py on a PVC machine to confirm the failure. Then trace the affected XPU backward kernels and their alertNotDeterministic usage, comparing the related fixes in issues #3117, #2512, #2516, and #2517. Done means the 25 listed tests match CUDA behavior without unexpected nondeterminism errors in default mode.

Written by the indexing model from the issue text.