Using sharding to parallelize indexing

**Is your feature request related to a problem? Please describe.**
Creating an elasticsearch index on large dataset could be quite long and cannot be parallelized on shard (the index creation is colliding)

**Describe the solution you'd like**
When working on dataset shards, if an index already exists, its mapping should be checked and if compatible, the indexing process should continue with the shard data.

Additionally, at the end of the process, the `_indexes` dict should be send back to the original dataset object (from which the shards have been created) to allow to use the index for later filtering on the whole dataset.

**Describe alternatives you've considered**
Each dataset shard could created independent partial indices. then on the whole dataset level, indices should be all referred in `_indexes` dict and be used in querying through `get_nearest_examples()`. The drawback is that the scores will be computed independently on the partial indices leading to inconsistent values for most scoring based on corpus level statistics (tf/idf, BM25).

**Additional context**
The objectives is to parallelize the index creation to speed-up the process (ie surcharging the ES server which is fine to handle large load) while later enabling search on the whole dataset.