## Bug Description

When a request arrives at LiteLLM's Anthropic pass-through endpoint (`/anthropic/v1/messages`) and is routed to a Claude Haiku 4.5 model backed by the `azure_ai/` provider, `output_config.effort` is forwarded to Azure AI Foundry unchanged. Azure rejects it with:

``` 400 {"type":"invalid_request_error","message":"This model does not support the effort parameter."} ```

This affects all callers using the Anthropic SDK pointed at a LiteLLM proxy — for example, **Claude Code** with `ANTHROPIC_BASE_URL` set to a LiteLLM proxy endpoint.

## Why Existing Workarounds Fail

`additional_drop_params: ["output_config"]` and `drop_params: true` operate on the OpenAI→provider translation layer. When the request enters via `/anthropic/v1/messages` and routes to an `azure_ai/` model, LiteLLM translates Anthropic→Azure AI format — but `additional_drop_params` is **not** applied during that translation step. The parameter survives into the outbound request.

## Existing Precedent

PR #29585 fixed the identical issue for Vertex AI: it strips `output_config.effort` for Vertex Claude models that reject it. The Azure AI provider path has no equivalent guard.

From `litellm/llms/vert…