Default truncation to second for text similarity

[davidkyle]

NLP models have 3 truncation settings: FIRST, SECOND and NONE

FIRST means truncate the first input. In most cases there is only 1 input (e.g for text embeddings) so this is a sensible default.
SECOND means truncate the second input. Task types with 2 inputs are extractive question answering where the question is one input and the context the other. Text Similarity takes has 2 inputs.
NONE means don't truncate and window the input.

For text similarity the first input is usually the shorter input, for example it might be the query text in a rerank operation. In this situation it is better to truncate the second input. This change makes that the default.

[pquentin]

Thanks! LGTM. Just need to remove the extra print.