[Bug Report]

**Extract from the workbook "machine_learning_workflow_abalone.ipynb"**
When adding tags in the following estimator :

mes_tags = [{'key': 'cart', 'value': 'dataengineering'}]

xgb = sagemaker.estimator.Estimator(
image_uris.retrieve("xgboost", region, "1.2-1"),
sagemaker_execution_role,
train_instance_count=1,
train_instance_type="ml.m4.4xlarge",
train_volume_size=5,
output_path=bucket_path + "/" + prefix + "/single-xgboost",
base_job_name=base_job_name,
**tags=mes_tags,**
sagemaker_session=session,
)
No error when creating the sagemaker.estimator object

**The workflow creation fails**
When running the command (later in the notebook):
workflow.create()
I got the exception :
"InvalidDefinition: An error occurred (InvalidDefinition) when calling the CreateStateMachine operation: Invalid State Machine Definition: 'SCHEMA_VALIDATION_FAILED: The field "key" is not supported by Step Functions at /States/Train Step/Parameters"

Which is clearly related to the tags I previously added.
Apparently, adding tags to a Sagemaker estimator in the training step does not seem to be supported by the current version of the SDK.

**To reproduce**
You can comment "tags=mes_tags" and re-rerun the notebook and the state machine is created without any errors.

**Logs**
Only the stack trace in the notebook

Start with machine_learning_workflow_abalone.ipynb and reproduce the failure with tags=mes_tags on the SageMaker estimator, then compare it with the version that omits tags. Trace the training step used by workflow.create() and verify that the resulting state-machine definition accepts the estimator configuration. Done means the workflow is created successfully while the requested tags are handled correctly.

Written by the indexing model from the issue text.