feat: Adding RestrictedFeatures Support to the Python Frontend Bindings

[KrishnanPrash]

What does the PR do?

This PR provides Restricted Features support to the python frontend bindings, a.k.a tritonfrontend.
This PR introduces 3 new classes to the tritonfrontend package:

1. Feature: 1-to-1 mapping of RestrictedCategory Enum
2. FeatureGroup: Pydantic @dataclass that stores (key, value, features) information while performing type validation.
3. RestrictedFeatures: Store collections of FeatureGroup instances, apply an additional layer of validation to check for collisions (ensuring that each feature belongs to only one group), and serialize the data into a JSON string.

With these changes, we can pass a RestrictedFeatures to the KServeHttp and KServeGrpc frontends, which looks along the lines of:

>>> import tritonserver
>>> from tritonfrontend import KServeHttp, KServeGrpc
>>> from tritonfrontend import RestrictedFeatures
>>> server = tritonserver.Server(...).start(wait_until_ready=True)

>>> rf = RestrictedFeatures(...)

>>> http_options = KServeHttp.Options(restricted_features=rf)
>>> http_service = KServeHttp(server, http_options)
>>> http_service.start()

>>> grpc_options = KServeGrpc.Options(restricted_features=rf)
>>> grpc_service = KServeGrpc(server, grpc_options)
>>> grpc_service.start()

A RestrictedFeatures instance can be created with:

>>> from tritonfrontend import Feature, FeatureGroup, RestrictedFeatures
>>> admin_group = FeatureGroup(key="admin-key", value="admin-value", features=[Feature.HEALTH, Feature.METADATA])
>>> infer_group = FeatureGroup("infer-key", "infer-value", [Feature.INFERENCE])

>>> rf = RestrictedFeatures([admin_group, infer_group])

>>> model_group = FeatureGroup("model-key", "model-value", [Feature.MODEL_REPOSITORY])
>>> rf.add_feature_group(model_group)

>>> rf.create_feature_group("stat-key", "stat-value", [Feature.STATISTICS])
>>> rf
[  
   {"key": "admin-key", "value": "admin-value", "features": ["health", "metadata"]}, 
   {"key": "infer-key", "value": "infer-value", "features": ["inference"]}, 
   {"key": "model-key", "value": "model-value", "features": ["model-repository"]}, 
   {"key": "stat-key", "value": "stat-value", "features": ["statistics"]}
]

Where should the reviewer start?

Take a look at api/_restricted_features.py and tritonfrontend.h

Test plan:

In L0_python_api/test_kserve.py added:

• test_correct_rf_parameters() and test_wrong_rf_parameters(): Tests for valid/invalid arguments passed/selected for Feature, FeatureGroup and RestrictedFeature
  To test if restricted features are working correctly with the frontends:
• test_restricted_features(): Restricts inference. Then, sends an inference request with a correct and incorrect header to verify proper functionality.

CI Pipeline ID: 20189649

Background

For more information about restricted features support in Triton, please take a look at this: [Link]

Additonal Changes Made in this PR

• Cleaned up includes in tritonfrontend.h and tritonfrontend_pybind.cc.
• Added support to @handle_triton_error for non-void functions.
• Added support to pass headers in send_and_test_inference_identity() so that we can re-use utility function for restricted features testing.

[KrishnanPrash]

Investigating why @field_validator is incompatible with @handle_triton_server nested underneath. To unblock, directly raised tritonserver equivalent error.

[GuanLuo]

Is this the pytest way to set test suit level parameters? Can you point me to the guide on that?

[GuanLuo]

test for:

1. multiple feature groups
2. overwrite feature groups
3. remove feature groups
4. duplicate feature in different groups

[GuanLuo]

what is rest_data?