[Feature Request] Cross entropy loss forward and backward functions

[rfurko-tt]

Is your feature request related to a problem? Please describe. To train a classification model we can do it in two ways:

• calculate log_softmax and pass it to NLLLoss (moreh_nll_loss is available)
• pass logits to cross entropy loss

Cross entropy will reduce amount of programs that we need to execute and improve performance (fused op).

Describe the solution you'd like Introduce cross_entropy_loss and cross_entropy_loss_backward that takes logits and targets (indices, int32_t or uint32_t)

Describe alternatives you've considered

• Use of existing NLLLoss
• Use of cross entropy with one-hot encoding implemented as a composite op

[dmakoviichuk-tt]

@ayerofieiev-tt moving it to you because I think Brian is the only person who can implement it fast because of his experience with embedding_bw