Add continual masking and trajectories

[GuZZ1119]

Summary

Adds the validation evidence and class-incremental evaluation guard requested after the continual-learning review.

• masks future, unseen classifier logits during continual validation while retaining all classes seen so far
• leaves training, fixed 100-class heads, and all existing held-out test outputs unchanged
• adds an offline trajectory plotter for saved training_history.csv files, with task boundaries and loss/Top-1 panels
• adds a dependency-light trajectory smoke check and CI coverage

Validation

• full manifest, continual-task, no-replay, replay, result-analysis, trajectory, and Grad-CAM smoke suite
• Python compilation
• CPU future-logit mask behaviour check
• synthetic PNG generation from saved-history-shaped CSV input
• git diff --check