[Tracker] Improve cuml.accel proxy layer implementation

[csadorf]

Current Implementation Limitations

The current proxy layer implementation (estimator_proxy.py) has architectural limitations that affect maintainability and extensibility:

1. Inheritance vs. Composition

The current design inherits from accelerated components rather than using composition:

class ProxyEstimator(class_b):  # Inherits from accelerated class
    def __init__(self, *args, **kwargs):
        self._cpu_model_class = original_class_a  # Stores original as reference

This creates tight coupling where:

• Accelerated components must handle translation logic
• Changes to either component affect the other
• Adding new acceleration paths requires modifying existing code

2. Early Parameter Translation

Parameter translation occurs at construction time rather than operation time:

def __init__(self, *args, **kwargs):
    translated_kwargs, self._gpuaccel = self._hyperparam_translator(**kwargs)
    super().__init__(*args, **translated_kwargs)

Issues:

• Parameters are translated before knowing runtime context
• Multiple translations may occur (original → accelerated → cpu)

3. Dispatch Limitations

Current dispatching is limited by:

• Limited support for branching at runtime
• Tight coupling with parameter translation

Requirements

P0 (Critical)

1. API Compatibility

   • Proxied objects must match original API behavior
   • Public interface (non _-prefixed) must be identical
   • Must preserve original error handling and warnings

2. Context-Aware Dispatch

   • Dispatch decision should consider the full context at operation time (fit(), predict(), etc.)
   • Consider full context:
     • Input data properties (type, shape, memory location)
     • Available hardware
     • Runtime state
     • Hyperparameters

P1 (Important)

1. Type System Integration

   # Should work correctly:
   isinstance(proxy_estimator, OriginalEstimator)  # True

2. Meta-Estimator Support

   • Support nested estimators
   • Preserve acceleration capabilities in pipelines
   • Handle parameter grids correctly

P2 (Nice-to-have)

1. Serialization

   # Should work with original library:
   import pickle
   loaded = pickle.loads(pickle.dumps(proxy_estimator))