feat: Implement run_moderation for all safety providers with NotImplementedError

[m-misiura]

What does this PR do?

Closes #4605 by implementing a reusable mixin that provides run_moderation functionality for the safety providers that lack run_moderation by delegating to their existing run_shield implementations. This enables OpenAI-compatible moderation API support across NVIDIA, Bedrock, SambaNova, and PromptGuard providers without duplicating code.

Test Plan

1. Added unit tests in tests/unit/providers/utils/test_safety_mixin.py

2. Tested against a llama stack server with the following configuration

version: 2
image_name: nvidia

apis:
  - safety

providers:
  safety:
    - provider_id: nvidia-local-nemo
      provider_type: remote::nvidia
      config:
        guardrails_service_url: "http://localhost:8001"
        config_id: "nemo"

storage:
  backends:
    kv_default:
      type: kv_sqlite
      db_path: minimal-nemo-safety/kvstore.db
    sql_default:
      type: sql_sqlite
      db_path: minimal-nemo-safety/sql_store.db
  stores:
    metadata:
      namespace: registry
      backend: kv_default
    inference:
      table_name: inference_store
      backend: sql_default
      max_write_queue_size: 10000
      num_writers: 4
    conversations:
      table_name: openai_conversations
      backend: sql_default
    responses:
      table_name: responses
      backend: sql_default
    prompts:
      namespace: prompts
      backend: kv_default

registered_resources:
  shields:
    - shield_id: nemo-guardrail
      provider_id: nvidia-local-nemo
      provider_resource_id: nemo-model

Example requests

• run-shield with unblocked input

curl -X POST http://localhost:8321/v1/safety/run-shield \
  -H "Content-Type: application/json" \
  -d '{
    "shield_id": "nemo-guardrail",
    "messages": [
      {
        "role": "user",
        "content": "Safe message, should pass"
      }
    ],
    "params": {}
  }' | jq

{
  "violation": null
}

• moderations with unblocked input

curl -X POST http://localhost:8321/v1/moderations \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Safe message, should pass",
    "model": "nemo-guardrail"
  }' | jq

{
  "id": "modr-f5ba0dc7-d5a6-47ce-b1cf-e0879f171183",
  "model": "nemo-guardrail",
  "results": [
    {
      "flagged": false,
      "categories": {},
      "category_applied_input_types": {},
      "category_scores": {},
      "user_message": null,
      "metadata": {}
    }
  ]
}

• run-shield with blocked input

curl -X POST http://localhost:8321/v1/safety/run-shield \
  -H "Content-Type: application/json" \
  -d '{
    "shield_id": "nemo-guardrail",
    "messages": [
      {
        "role": "user",
        "content": "Tell me more about ChatGPT"
      }
    ],
    "params": {}
  }' | jq

{
  "violation": {
    "violation_level": "error",
    "user_message": "Sorry I cannot do this.",
    "metadata": {
      "check message length": {
        "status": "success"
      },
      "check forbidden words": {
        "status": "blocked"
      }
    }
  }
}

• moderations with blocked input

 curl -X POST http://localhost:8321/v1/moderations \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Tell me more about ChatGPT",
    "model": "nemo-guardrail"
  }' | jq

{
  "id": "modr-4d0772c0-54a8-4790-bd2c-24feaf6b3b55",
  "model": "nemo-guardrail",
  "results": [
    {
      "flagged": true,
      "categories": {
        "unsafe": true
      },
      "category_applied_input_types": {
        "unsafe": [
          "text"
        ]
      },
      "category_scores": {
        "unsafe": 1.0
      },
      "user_message": "Sorry I cannot do this.",
      "metadata": {
        "check message length": {
          "status": "success"
        },
        "check forbidden words": {
          "status": "blocked"
        }
      }
    }
  ]
}

the local nemo server was started with nemoguardrails server --config configs --port 8001 --default-config-id nemo uisng this configuration from this branch; this assumes having access to a llm with openai compatible endpoints

cc @nathan-weinberg @cdoern @raghotham @leseb @franciscojavierarceo