[ML] Support for type aggregate_metric_double in the datafeed

[valeriy42]

At the moment if the datafeed encounters a field with type aggregate_metric_double when using scroll API it fails with the message that

type [aggregate_metric_double] and cannot be used in a datafeed without aggregations

However, the anomaly detection job seems to be able to handle the indices patterns with the mix of aggregate_metric_double and double types for the same field without any problems.

Since downsampling data streams per ILM policy is a commonly used option, the field type's support in anomaly detection must also be extended.

[valeriy42]

I created a minimal example with an index pattern with mixed types and it works.

Due to downsampling, the variance is not modelled as well as without downsampling, but technically, this should not be a showstopper.

Methodologically, the bucket_span should be a multiple of the downsampling's fixed_interval for the modelling to work meaningfully.

[elasticsearchmachine]

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