SQL: normalized keywords shouldn't be allowed for groupings and sorting

[astefan]

A normalized keyword field can potentially change the original value of a field, but SQL works with exact values. Probably as part of #34718, the restriction was lifted, as field_caps API can tell if a field is aggregatable/searchable, but not if it's normalized or a simple keyword.

As such, in the following scenario:

{
  "settings": {
    "analysis": {
      "normalizer": {
        "my_normalizer": {
          "type": "custom",
          "filter": [
            "lowercase",
            "asciifolding"
          ]
        }
      }
    }
  },
  "mappings": {
    "test": {
      "properties": {
        "user": {
          "type": "text",
          "fields": {
            "normalized": {
              "type": "keyword",
              "normalizer": "my_normalizer"
            },
            "keyword": {
              "type": "keyword"
            }
          }
        },
        "user2": {
          "type": "keyword",
          "normalizer": "my_normalizer"
        }
      }
    }
  }
}

With a query like { "query" : "select user from test group by user" } one gets back:

{
    "error": {
        "root_cause": [
            {
                "type": "mapping_exception",
                "reason": "Multiple exact keyword candidates available for [user]; specify which one to use"
            }
        ],
        "type": "mapping_exception",
        "reason": "Multiple exact keyword candidates available for [user]; specify which one to use"
    },
    "status": 400
}

And for { "query" : "select user2 from test group by user2" } there are results returned where, in fact, ES SQL should return an error.

[elasticmachine]

Pinging @elastic/es-search-aggs