[TORCH] Add Kullback-Leibler divergence loss support

[zahidwx]

This PR takes care of #4203.

• e2e support of aten.kl_div op supporting all reduction modes (mean, sum, batchmean, none)
• reduction: batchmean requires special handling by calling op with sum and then dividing it by input batch_size.

Following tests are failing and are marked either in expected failures or crashing set

• KlDivLossModule_batchmean_reduction_basic
  • config=linalg | RuntimeError: attribute lookup is not defined on builtin | LINALG_XFAIL_SET
• KlDivLossModule_default_basic | KlDivLossModule_reduction_is_none_basic | KlDivLossModule_mean_reduction_basic | KlDivLossModule_sum_reduction_basic | KlDivLossModule_batchmean_reduction_basic
  • config=torchdynamo | error: failed to legalize operation 'torch.aten.xlogy.Tensor' | TORCHDYNAMO_CRASHING_SET
• KlDivLossModule_batchmean_reduction_basic
  • config=onnx | Error: RuntimeError: aten::div() Expected a value of type 'number' for argument 'other' but instead found type 'Tensor' Position: 1
    Value: tensor(1)
    Declaration: aten::div.Scalar(Tensor self, Scalar other) -> Tensor
    Cast error details: Cannot cast tensor(1) to number | ONNX_XFAIL_SET
• KlDivLossModule_default_basic | KlDivLossModule_reduction_is_none_basic | KlDivLossModule_reduction_is_none_log_target_is_true_basic | KlDivLossModule_mean_reduction_basic | KlDivLossModule_sum_reduction_basic | KlDivLossModule_batchmean_reduction_basic
  • config=onnx_tosa | Error: error: failed to legalize operation 'torch.aten.size.int' that was explicitly marked illegal | ONNX_TOSA_XFAIL_SET

[zahidwx]

Can someone review this please? @vivekkhandelwal1 @penguin-wwy @AmosLewis @rsuderman