feat: add window training modes for signal/background regions

[kratsg]

Summary

• Implements modes 3 and 4 from issue train cuts that select regions, also regions to reject #7
• Adds automatic window learning to keep signal events (mode 3: window_keep)
• Adds dual network approach to learn rejection windows for background events (mode 4: window_reject)
• Maintains full backward compatibility with existing modes 1 and 2

Implementation Details

Mode 3: Window Keep (mode="window_keep")

• Automatically duplicates each input feature to learn both upper and lower bounds
• Uses fixed weights [+1, -1] pairs with learnable biases
• Learns windows where signal events are kept: x1 < x < x2
• Eliminates need to manually pass inputs twice with different signs
• Returns cuts as (original_features, 2) shaped tensor

Mode 4: Window Reject (mode="window_reject")

• Implements dual network architecture with acceptance and rejection networks
• Acceptance network learns to keep signal (normal targets)
• Rejection network learns windows to reject background
• Final output: accept_prob * (1.0 - reject_prob)
• Learns regions where background is rejected: x < x1 || x > x2

Architecture Changes

• Enhanced OneToOneLinear with mode parameter and input expansion logic
• Added DualNetworkLinear class for dual network approach
• Updated EfficiencyScanNetwork to support different training modes
• Modified loss functions to handle window cuts regularization properly
• Preserved cabin's core philosophy of uncorrelated cuts in network

Test Plan

• Unit tests for mode 3 (window_keep) functionality
• Unit tests for mode 4 (window_reject) functionality
• Backward compatibility tests for existing modes 1 and 2
• All existing tests pass without modification
• Cut shape validation for windowed modes
• Training step validation for all modes

Usage Examples

# Mode 3: Learn windows to keep signal
efficnet = cabin.EfficiencyScanNetwork(
    features=10, 
    effics=[0.5, 0.7], 
    mode="window_keep"
)

# Mode 4: Learn rejection windows for background
efficnet = cabin.EfficiencyScanNetwork(
    features=10, 
    effics=[0.5, 0.7], 
    mode="window_reject"
)

🤖 Generated with Claude Code