Dependency injection for model deployments

[yifanmai]

This allows using ModelDeploymentRegistry with clients that do not have a api_key parameter.

Example:

helm-run --run-specs boolq:model=simple/my-simple-model --suite debug -m 5 --model-deployment-paths model_deployments_simple.yaml --model-metadata models.yaml

models.yaml:

models:
  - name: simple/my-simple-model

model_deployments_simple.yaml:

model_deployments:
  - name: simple/my-simple-model
    model_name: simple/my-simple-model
    tokenizer_name: "huggingface/gpt2"
    max_sequence_length: 2048
    client_spec:
      class_name: "helm.proxy.clients.simple_client.SimpleClient"
      args: {}

[percyliang]

Elaborate on documentation of additional_args

[percyliang]

Hmm...I think it might be cleaner / more modular to have a function that takes a spec and updates the spec with additional_args, so that the create_object function can stay the same.

[percyliang]

I guess I'm confused what injectors are supposed to do...

[yifanmai]

Closing because this is addressed by #1861.