feat: Cosby PTF prior for K_D, combined optimizers, attribute column selection

[taddyb]

Summary

• K_D reparameterized as Cosby PTF + learned delta: Instead of learning K_D directly from bounds, K_D is now computed as 10^(log10_Ks_cosby + delta) where delta is the KAN-learned correction. This anchors hydraulic conductivity to the Cosby et al. (1984) pedotransfer function using sand/clay percentages, with the neural network learning a log-space correction term (K_D_delta with bounds [-2, 2]).
• Unified attribute column selection: KAN and LSTM now each receive only their own input variables via a new select_columns() utility, drawn from a union attribute tensor. RoutingDataclass carries attribute_names for name-based column lookup.
• Switched KAN optimizer to Adadelta: Both KAN and LSTM now use Adadelta (no manual LR schedule). Removed learning_rate from ExperimentConfig and the resolve_learning_rate helper.
• Dead code cleanup: Removed resolve_learning_rate() from scripts_utils.py and its tests.
• Docs updated: Training guide and future_work docs reflect the new dual-network architecture and Cosby PTF approach.

Test plan

• Verify leakance tests pass (tests/routing/test_leakance.py)
• Verify routing utils tests pass (tests/routing/test_utils.py)
• Verify config validation tests pass (tests/routing/test_leakance.py::TestLeakanceConfigValidation)
• Verify scripts_utils tests pass without resolve_learning_rate tests
• Run training on MERIT/Lynker datasets to confirm end-to-end compatibility

🤖 Generated with Claude Code