Sentecnes shared by NLI problems

NLI datasets usually contain NLI problems `` that share sentences.
For example, this is the case for the SICK, FraCaS, and SNLI datasets.

```
SICK-56 SICK-58
P: Three kids are sitting in the leaves P: Children in red shirts are playing in the leaves
H: Three kids are jumping in the leaves H: Three kids are sitting in the leaves
```

The question is whether to distinguish or not such identical sentences from each other based on their corresponding NLI problem IDs and the sentence position (e.g., `56P` vs `58H`).

Pros and cons of not distinguishing identical sentences.
➕ annotation economy: no need to annotate (e.g., correct LLFs or parse trees) multi-occurring sentences several times;
➕ complicates the annotation workflow;
➖ not accounting for ambiguity: If a sentence is ambiguous (e.g., three kids and leaves), and different meanings are relevant for different NLI problems.

For simplicity, I suggest not distinguishing sentences based on their involvement in other NLI problems.
The annotation process will show us the trade-off of this decision.
Later, we can always switch to the distinguishing option more smoothly than in the opposite direction.