Fixed alignment upsample index offset

[Jamy-L]

The indexes computed are applied to paddedPreviousAlignments which is only offseted by 1 (and not 2) pixels with regard to PreviousAlignments. This is not such a big deal, but this offset changes 2 of the 3 candidates alignments by taking them much further than what is indicated in the IPOL article. It can easily be seen on toy-cases that the returned indexes are not coherent.