[FEA] Improve the processing of `Targets` in ItemPrediction tasks

[sararb]

🚀 Feature request

Currently, all ItemPrediction tasks transform the targets tensor to a one-hot representation (as required by the RankingMetrics or SampledSoftmax contrastive learning).

This Feature request is related to setting the categorical encoded representation of the target (i.e removing the one-hot transformation block) as the default representation of the ItemPredictionTask and setting the one-hot conversion within the top-k evaluation metric class.