[ML] Ad-hoc Data Views for testing models in Trained Models UI

[darnautov]

Currently to retrieve a list of text fields from the source data to test a model output we rely on Kibana Data Views. For users who mostly use the Elasticsearch API to ingest data and are not familiar with the concept of data views, it might be an unnecessary friction point. We can improve the UX by providing ad-hoc data views either for

1. all indices containing text fields

or

2. narrowing down the list to indices that utilize ingest pipelines with inference processors of the testing model

[elasticmachine]

Pinging @elastic/ml-ui (:ml)

[serenachou]

@joemcelroy - how do you flag that semantic_text or dense_vector is being used for Playground? Should we use similar logic to what's proposed here for #2 option?

@darnautov we could potentially also key off of semantic_text field mapping being present?

[joemcelroy]

In playground we perform a field-caps API request and a single hit query to gather the schema and the model_id that was used for inference.

Im expecting when semantic_text lands, the field type will be semantic or similar which the field caps API will help on.

We don't need to get the pipeline processor as I get the model_id from the hit itself but its another option to get the model_id.