benchmark: complete robust Huber Bisquare Fair backend matrix

[TheHiddenObserver]

Goal

Add current, statistically aligned NumPy/CuPy/Torch and external-reference benchmark coverage for robust Huber, Bisquare, and Fair losses.

Required work

• define identical datasets, scale identities, solver/tolerance settings, and timing boundaries;
• synchronize CuPy and Torch timings;
• distinguish fit timing, convergence, accuracy, and inference where supported;
• preserve failure and convergence counts;
• produce a structured source, parser tests, manifest registration, generated assets, and coverage-matrix update.

No missing measurements may be inferred from existing summaries.

Parent: #100