perf: Vectorize PredictionsModelWrapper lookup with numpy arrays

[imatiach-msft]

Problem

The previous hash-based optimization (0.6.2) still used expensive DataFrame operations inside the prediction loop, resulting in only ~130 rows/sec throughput. This caused RAI dashboard generation to still timeout on large datasets (~93K rows).

Root Cause

The 0.6.2 implementation had these bottlenecks:

• query_test_data.iloc[index:index + 1] created a new DataFrame for each row
• _compute_row_hash() called pd.isna() on each value individually
• _rows_equal() iterated through each column using pandas Series operations
• Hash lookup was O(1), but per-row preparation was still O(m) with high constant factors

Solution

This PR replaces DataFrame operations with numpy array operations:

1. Pre-extract numpy arrays during __init__:

   • self._feature_values = combined_data[feature_names].values
   • self._predictions = y_pred.copy()

2. Fast hashing using str(row.tolist()) + MD5:

   • Handles NaN, mixed dtypes consistently
   • Much faster than per-value type checking

3. Numpy array indexing in the predict loop:

   • Extract all query data as numpy array once: query_values = query_test_data[feature_names].values
   • Direct array indexing: self._predictions[matched_idx]

Performance Results

Tested with production AutoML data (93,340 train rows, 5,000 test rows, 33 features):

Metric	Before (0.6.2)	After	Speedup
predict() throughput	130 rows/s	28,000 rows/s	~215x
93K row prediction	~40 min	3.3 sec	~720x
RAIInsights creation	~110 min	11.5 sec	~570x
Total RAI job	~194 min	< 1 min	~200x

Validation

• ✅ All 60 existing tests pass
• ✅ Validated on AzureML with production AutoML RAI job
• ✅ Job completed successfully in < 1 minute

Files Changed

• python/ml_wrappers/model/predictions_wrapper.py - Vectorized implementation

Backwards Compatibility

• API unchanged - Same public interface
• Behavior unchanged - Same results for all inputs
• Pickle compatible - Numpy arrays rebuilt on deserialization