Initial checkpoint

This snapshot of the code is working end-to-end. We can train models on any chosen languages, both in MTL and meta-learning setting. The evaluation can either be zero-shot (full) on the whole dataset, or meta-testing (meta) which is NOT on the whole dataset, but only on a sample (with size = k i.e. number of shots).