add branch level predictions to output

Summary
The graphkit predict command (and the --save_output flag on train/test) currently writes a single bus-level parquet file for the PowerFlow task. Branch-level quantities such as active/reactive power flows, thermal loading, angle differences, and violation flags are already computed within predict_step, but are not currently included in the saved outputs.
Exposing these branch-level results would enable downstream workflows such as branch-level uncertainty analysis, congestion assessment, and N-1 security studies directly from GraphKit outputs, without requiring users to reimplement branch flow calculations outside the framework.

Desired behaviour
Extend predict_step to return a two-table dictionary, following the pattern already used by the OPF task, allowing both bus-level and branch-level outputs to be written by the existing CLI logic.
This would produce:
• {grid_name}_predictions.parquet — bus-level results (unchanged)
• {grid_name}_branch_predictions.parquet — branch-level results (new)

Rationale
Branch-level metrics are often the quantities of interest for operational and planning studies, including:
• Thermal loading and overload analysis
• Congestion assessment
• N-1 security evaluation
• Probabilistic and uncertainty-based risk studies

Since these quantities are already computed during inference, making them available in the saved outputs would improve the usability of the framework for downstream analysis while requiring only a small extension to the existing output structure.

Start at predict_step for the PowerFlow task and compare its return structure with the OPF task, then trace the existing predict command and --save_output CLI logic. Extend the output path so bus-level predictions remain in {grid_name}_predictions.parquet and branch-level results are written to {grid_name}_branch_predictions.parquet.