[torchlib] Support class probability targets in aten::cross_entropy_loss

[Raashish Aggarwal (raashish1601)]

F.cross_entropy also accepts class probabilities as the target (same shape as the input, float dtype), which is common for soft labels, mixup and distillation. aten_cross_entropy_loss always lowered to SoftmaxCrossEntropyLoss, which only accepts integer class indices, so the export produced an invalid model:

x = torch.tensor([[2.0, 0.0, 0.0]]); t = torch.tensor([[0.7, 0.2, 0.1]])
F.cross_entropy(x, t)  # 0.8395
# onnxruntime: INVALID_GRAPH ... Type 'tensor(float)' of input parameter ... of operator (SoftmaxCrossEntropyLoss)

For a floating-point target this now follows PyTorch's cross_entropy_loss_prob_target: -sum(log_softmax(x) * target * weight) over the class dim, with label_smoothing applied to the target (target * (1 - eps) + eps / C). mean divides by the number of elements excluding the class dim, as PyTorch does for probability targets. Unbatched (C,) input is handled too. Integer targets keep the existing SoftmaxCrossEntropyLoss path, and the target annotation changes from IntType to TensorType.

Testing:

• Removed the xfail for non-int64 targets on nn.functional.cross_entropy. Those 20 samples fail on main and pass with the change. pytest tests/function_libs/torch_lib/ops_test.py -k cross_entropy passes locally (onnxruntime 1.23.0).
• I also compared torch.onnx.export(..., dynamo=True) + onnxruntime against eager PyTorch for probability targets on (C,), (N, C), (N, C, d1) and (N, C, d1, d2) inputs with every reduction, with and without weight, and with label_smoothing 0 and 0.2 (48 cases). All match within 1e-5.
• ruff check and format are clean on the changed files.

This touches the same function as #3069 (label_smoothing for index targets), but the two branches merge without conflicts and the combined tests pass.