Optimize tensor deduplication function

[mefor44]

During src.inference there is a step that deduplicates the same semantic IDS of items. Items with the same semantic ID get an additional id that is an int (starting from 0). Items without duplicates get additional 0 token. The function deduplicate_rows_in_tensor from src.utils.tensor_utils.py is super slow when the number of items is large (in millions). I provide much more efficient implementations of this function