[ML] Track token positions and use source string to tag NER entities

[davidkyle]

By recording the position of the tokens in the original source string the entity labels can correctly constructed.

• Case is preserved
• Accents and other characters stripped during normalisation are preserved
• Fixes bugs with extra spaces or punctuation

The input "My name is Benjamin Trent, I work at Acme Inc.." now correctly groups Acme Inc.

{
  "predicted_value" : "My name is [Benjamin Trent](PER&Benjamin+Trent), I work at [Acme Inc.](ORG&Acme+Inc.).",
  "entities" : [
    {
      "entity" : "Benjamin Trent",
      "class_name" : "PER",
      "class_probability" : 0.9994054708878661,
      "start_pos" : 11,
      "end_pos" : 25
    },
    {
      "entity" : "Acme Inc.",
      "class_name" : "ORG",
      "class_probability" : 0.818787531972483,
      "start_pos" : 37,
      "end_pos" : 46
    }
  ]
}

Monsieur Gillenormand is correctly grouped : "M. Gillenormand, who had risen betimes like all old men in good health, had heard his entrance"

{
  "predicted_value" : "[M. Gillenormand](PER&M.+Gillenormand), who had risen betimes like all old men in good health, had heard his entrance",
  "entities" : [
    {
      "entity" : "M. Gillenormand",
      "class_name" : "PER",
      "class_probability" : 0.9861165998812788,
      "start_pos" : 0,
      "end_pos" : 15
    }
  ]
}

[elasticmachine]

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

[davidkyle]

run elasticsearch-ci/part-2