Calibration enhancements - Group weighting

[baogorek]

Closes #30

Summary

This PR enhances the calibration module with group-wise loss averaging and improved training output for better interpretability.

Key Changes

1. Group-wise Loss Averaging: Added support for balancing loss contributions from targets with different cardinalities. This is particularly useful when calibrating with mixed targets like:

   • National-level targets (singletons with large values)
   • State-level age distributions (groups with many bins)

   The new target_groups parameter ensures each group contributes equally to the loss, preventing high-cardinality groups from dominating the optimization.

2. Improved Training Output: Enhanced verbose output to show:

   • Relative error percentages instead of raw loss values
   • Sparsity statistics (active weights and percentage)
   • Group-averaged errors when using group-wise averaging
   • More interpretable progress monitoring

3. Simplified Active Weight Detection: Removed the threshold parameter from get_active_weights() - now simply checks for weights > 0.

Example Usage

# Group targets for balanced optimization
target_groups = np.array([
    0, 1, 2,      # 3 national targets (singletons)
    *[3]*18,      # 18 age bins for state 1
    *[4]*18       # 18 age bins for state 2
])

model.fit(
    M, y,
    target_groups=target_groups,  # Balance group contributions
    verbose=True
)

Test Coverage

• Added comprehensive tests for group-wise averaging
• Tested edge cases (single group, all singletons, unbalanced groups)
• Verified that grouping improves balance between different target types

Output Example

Before:

Epoch  100: loss=0.1234, data_loss=0.1000, mse=0.0500, l0=25.00, active=2500, mean_weight=1.234

After:

Epoch  100: mean_group_loss=2.5%, max_error=8.3%, total_loss=0.123, active=2500/10000 (75.0% sparse)

The new output format is more intuitive and directly shows the calibration quality as percentages.