[ENHANCEMENT] Avoid un-necessary predictions in ingest processors

[mingshl]

Is your feature request related to a problem?
When re-indexing, text_embedding processors and ml_inference processors will run the prediction again even when the inference fields are already existed in the document. text_embedding processors will over-write the the inference field, while ml_inference will not write to the inference field, then throw exception or skip writing to the document.

What solution would you like?
In the ingest processor that used ml inference, we should check the model output field name, for example, text_embedding field exists in the document before the prediction tasks run. If the field already exists, skip the predictions.

In this case, when re-index happens, it won't run the prediction tasks again if the field is existed.

What alternatives have you considered?
Welcome any other suggestions.

[IanMenendez]

Maybe it would be nice to add a parameter on ml ingest processors called overwrite: Option[Boolean] if true it overwrites current embeddings upon reindexing if false it does not overwrite

It is nice to still have the option to overwrite in case we change the ML model to one that outputs different embeddings

[ylwu-amzn]

Added in this PR #2508