(sagemaker): Missing support for "AsyncInferenceConfig" in `create_endpoint_config` and `describe_endpoint_config`

[bweigel]

The SageMaker client does not fully support handling of the "AsyncInferenceConfig" parameter for creating endpoint configs.

While investigating a test failure in another project (which relies on async inference), I found that create_endpoint_config did not actually take "AsyncInferenceConfig" into account, since describe_endpoint_config did not return it. This missing functionality caused tests using async inference to break unexpectedly.

Steps to Reproduce

Details

import boto3
from moto import mock_aws
import pytest

@pytest.fixture
def mock_sagemaker():
    with mock_aws():
        yield boto3.client("sagemaker", region_name="eu-central-1")

@pytest.mark.parametrize(
    "prefix, endpoint_cfg, expected_result",
    [
        (
            "async", {
                "AsyncInferenceConfig": {
                    "ClientConfig": {"MaxConcurrentInvocationsPerInstance": 3},
                    "OutputConfig": {"S3OutputPath": "s3://output-bucket", "NotificationConfig": {}},
                }
            },
            True,
        ),
        ("", {}, False),
    ],
)
def test_is_async_endpoint(mock_sagemaker, prefix, endpoint_cfg, expected_result):
    ## given
    _MODEL_NAME = f"{prefix}test"
    _ENDPOINT_NAME = f"{prefix}test_endpoint_name"
    sm = mock_sagemaker
    sm.create_model(
        ModelName=_MODEL_NAME,
        PrimaryContainer={
            "Image": "test_image",
            "ModelDataUrl": f"s3://test_bucket/model.zip",
        },
    )
    sm.create_endpoint_config(
        EndpointConfigName=_ENDPOINT_NAME,
        ProductionVariants=[
            {
                "VariantName": "AllTraffic",
                "ModelName": _MODEL_NAME,
                "InitialInstanceCount": 1,
                "InstanceType": "ml.m5.large",
            }
        ],
        **endpoint_cfg,
    )
    sm.create_endpoint(
        EndpointName=_ENDPOINT_NAME,
        EndpointConfigName=_ENDPOINT_NAME,
    )
    ## when & then
    config_name = sm.describe_endpoint(EndpointName=_ENDPOINT_NAME)['EndpointConfigName']
    endpoint_config = sm.describe_endpoint_config(EndpointConfigName=config_name)
    assert ("AsyncInferenceConfig" in endpoint_config) is expected_result

Expected Behavior

• When specifying "AsyncInferenceConfig" in create_endpoint_config, Moto should store this parameter and return it when describe_endpoint_config is called.

Actual Behavior

• Moto ignores the "AsyncInferenceConfig" parameter in create_endpoint_config and does not surface this info in describe_endpoint_config, leading to test failures for async endpoint usage.

Proposed Solution

• Update FakeEndpointConfig.__init__ & SageMakerModelBackend.create_endpoint_config to include "AsyncInferenceConfig" params
• Update SageMakerResponse.create_endpoint_config to include "AsyncInferenceConfig" params