[NLP] Support the different mask tokens used by NLP models for Fill Mask

[maxhniebergall]

This change adds a mask_token field to the GET _ml/trained_models API, as an enhancement to support kibana#159577.

The mask_token field reveals the particular mask token for needed each model deployed. Below is an example inference_config component of the API response, with the new mask_token field within fill_mask.

      "inference_config": {
        "fill_mask": {
          "vocabulary": {
            "index": ".ml-inference-native-000001"
          },
          "tokenization": {
            "roberta": {
              "do_lower_case": false,
              "with_special_tokens": true,
              "max_sequence_length": 512,
              "truncate": "first",
              "span": -1,
              "add_prefix_space": false
            }
          },
          "num_top_classes": 5,
          "mask_token": "<mask>"
        }
      },

[elasticsearchmachine]

Pinging @elastic/ml-core (Team:ML)

[elasticsearchmachine]

Hi @maxhniebergall, I've created a changelog YAML for you.

[droberts195]

This largely looks good now.

One more thing that needs changing is the current reference docs. We need a new entry for mask_token immediately before this line:

elasticsearch/docs/reference/ml/trained-models/apis/get-trained-models.asciidoc

Line 177 in dc8a946

`tokenization`::::

[elasticsearchmachine]

Hi @maxhniebergall, I've updated the changelog YAML for you.

[maxhniebergall]

@elasticsearchmachine run elasticsearch-ci/rest-compatibility

[davidkyle]

LGTM