[ML] Dynamically get of num allocations for ml node models

[davidkyle]

Reinstates the ability to dynamically report the number of allocations for ml node models in the Inference API.

Dynamically reading the number of allocations in use on GET is useful for the case where adaptive allocations is enabled and if the model deployment is managed through ml trained models. The num_allocations field in service_settings is updated with the current value.

GET _inference/elser_on_ml

{
  "endpoints": [
    {
      "inference_id": "elser_on_ml",
      "task_type": "sparse_embedding",
      "service": "elasticsearch",
      "service_settings": {
        "num_allocations": 3,            <-- this field is dynamic
        "num_threads": 1,
        "model_id": ".elser_model_2",
        "deployment_id": ".elser_model_2_for_me"
      },
      "chunking_settings": {
        "strategy": "sentence",
        "max_chunk_size": 250,
        "sentence_overlap": 1
      }
    }
  ]
}

The change was originally reverted in 2697f85 due to an error calling the GroupedActionListener with 0 request. That bug is fixed in cce77b9.

This RP also includes bug fixes forward ported from 8.16 added in #115095

[elasticsearchmachine]

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

[elasticsearchmachine]

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

[jan-elastic]

LGTM

[elasticsearchmachine]

💔 Backport failed

Status	Branch	Result
❌	8.x	Commit could not be cherrypicked due to conflicts

You can use sqren/backport to manually backport by running backport --upstream elastic/elasticsearch --pr 115233