[torchlib] Support label_smoothing in aten::cross_entropy_loss

[Raashish Aggarwal (raashish1601)]

aten_cross_entropy_loss accepted label_smoothing but ignored it, so models that use F.cross_entropy(..., label_smoothing=...) (or nn.CrossEntropyLoss(label_smoothing=...)) export without errors but compute a different loss:

x = torch.tensor([[2.0, 0.0, 0.0]]); t = torch.tensor([0])
F.cross_entropy(x, t, label_smoothing=0.1)  # 0.3729
# exported model returns 0.2395 (the loss without smoothing)

This PR follows PyTorch's cross_entropy_loss_label_smoothing. It takes the log_prob output of SoftmaxCrossEntropyLoss, sums -log_prob (weighted when weight is given) over the class dim, zeros it where target == ignore_index, and reduces it the same way as the nll term (for mean, divided by the summed weights or the count of non-ignored targets). The result is (1 - eps) * nll + eps / C * smooth. When label_smoothing == 0 the graph is unchanged.

Testing:

• The nn.functional.cross_entropy OpInfo samples in PyTorch don't use label_smoothing, so I added an ops.aten.cross_entropy_loss.label_smoothing OpInfo with index targets, all three reductions, with and without weight, and with an ignore_index that hits some targets (2D and 4D inputs). Its 24 subtests fail on main and pass with the change.
• pytest tests/function_libs/torch_lib/ops_test.py -k cross_entropy passes locally.
• I also compared torch.onnx.export(..., dynamo=True) + onnxruntime against eager PyTorch for 72 combinations (2D/3D/4D inputs, reductions, weight, ignore_index, label_smoothing 0 and 0.2). All match within 1e-5.
• ruff check and format are clean on the changed files.