Add task-combination ablation analysis for manuscript revision

[Caroline070590]

This pull request adds the full auxiliary-task combination ablation analysis used for the manuscript revision.

The analysis includes:

• all 15 non-empty combinations of the four LISBET auxiliary tasks;
• five matched random seeds per combination;
• 75 trained model configurations in total;
• the executed Slurm job definitions and experiment matrix;
• the training-summary plotting script;
• environment and provenance documentation.

The experiment used a window size of 200 frames, 75 training epochs, 1,600 training samples, batch size 32, and learning rate 1e-4.

Raw datasets, model weights, private cluster paths, and CalMS21 downstream decoding analyses are not included.