Temporal Phi KAN experiments with forcing readers and upgraded losses

[taddyb]

Summary

• Implement new temporal phi KAN with Qprime input at the daily scale
• Add forcing readers and date range support for data loading
• Add normalization values and remove in-place operations for training stability
• Upgrade loss functions with additional validation utilities
• Add bias correction integration tests

Changes

• scripts/train.py / scripts/train_and_test.py: Extended training and evaluation pipelines for the new KAN architecture
• src/ddr/nn/temporal_phi_kan.py: Refactored temporal phi KAN with Qprime daily-scale input
• src/ddr/io/readers.py / src/ddr/io/statistics.py: New forcing data readers and statistics computation
• src/ddr/validation/losses.py: Upgraded loss functions
• src/ddr/validation/configs.py / enums.py / utils.py: Extended configs and validation enums
• src/ddr/geodatazoo/dataclasses.py: Additional dataclass fields
• src/ddr/routing/mmc.py: Minor routing adjustments
• tests/: Updated KAN tests and added bias correction integration tests

Test plan

• Run existing unit tests (tests/nn/test_temporal_phi_kan.py)
• Run new integration tests (tests/routing/test_bias_correction_integration.py)
• Verify training script runs end-to-end with updated config

🤖 Generated with Claude Code