[ML] add new multi custom processor for data frame analytics and model inference

[benwtrent]

This adds the multi custom feature processor to data frame analytics and inference.

The multi processor allows custom processors to be chained together and use the outputs from one processor as the inputs to another.
Example:

{
    "multi_encoding": {
        "processors": [
            {
                "ngram_encoding" : {
                    "field": "foo",
                    "feature_prefix": "f",
                    "n_grams": [1],
                    "length": 2   
                }
            },
            {
                "one_hot_encoding": {
                    "field": "f.11", //OUTPUT from earlier ngram_encoding
                    "hot_map": {"a": "col_a"}
                }
            },
            {
                "one_hot_encoding": {
                    "field": "some_additional_doc_field", // Some other outside field
                    "hot_map": {"cat": "col_cat"}
                }
            }
        ]
    }
}

This definition has the required input fields of ["foo", "some_additional_doc_field"] and has the output fields of ["f.12", "col_a", "col_cat"]. f.12 is included as it is an output of ngram_encoding but is not used by any other processor in the array.

[elasticmachine]

Pinging @elastic/ml-core (:ml)

[dimitris-athanasiou]

Looks good! A few comments to get through.

[dimitris-athanasiou]

I do not understand why we need to setOrdered. I think I also do not understand what orderedModeCallback is all about, it seems to me it's just a way to set some additional property on the created objects. And I find it strange that the other version of declareNamedObjects that doesn't take an orderedModeCallback passes in one that throws if the value is an array given it's declareNamedObjects plural. Not really this PR's issue but I would like to clear this up. Could you help us with this @nik9000 please?

[benwtrent]

@dimitris-athanasiou a user can declare named objects in two ways:

• a hashmap keyed by the named object names (unordered)
• an array of objects (ordered)

If I did not pass this call back (and subsequently check that it was called in the builder), the user could pass an unordered hashmap, which we don't allow.

Example of unordered :

{
    "multi_encoding": {
        "processors":
            {
                "ngram_encoding" : {
                    "field": "foo",
                    "feature_prefix": "f",
                    "n_grams": [1],
                    "length": 2   
                },
                "one_hot_encoding": {
                    "field": "f.11", //OUTPUT from earlier ngram_encoding
                    "hot_map": {"a": "col_a"}
                }
        }
}

^ This is a valid named objects collection, but obviously, its order is not guaranteed. We need to throw in this case.

As for making the declareNamedObjects array parsing more clear, that is probably best for another PR.

[dimitris-athanasiou]

LGTM