### Description `litellm.completion(..., max_retries=N)` is only honored for OpenAI and Azure. For every other provider (Anthropic, Cohere, Bedrock, Vertex, etc.) the parameter is silently dropped. No retries happen and no error is raised.

### Reproduction ```python import litellm

# Works: OpenAI retries at the SDK layer litellm.completion(model="gpt-4o", messages=[...], max_retries=5)

# No effect: Anthropic never retries, param is ignored litellm.completion(model="anthropic/claude-3-5-sonnet-20241022", messages=[...], max_retries=5) ```

### Root cause In `litellm/utils.py` `get_optional_params`, `_check_valid_arg` explicitly skips `max_retries` for non-OpenAI providers (legacy TODO). OpenAI/Azure consume it via the SDK client constructor; other providers never receive it and have no retry transport configured.

### Expected behavior `max_retries` should be a supported, retry-honored parameter for all providers, consistent with OpenAI/Azure.