CLI: L1 loss and PSNR with Self-Supervised Loss for benchmark_algos command

[AhmetCanSolak]

This PR introduces a new result metric to Aydin CLI benchmark_algos command, which is PSNR with self-supervised loss used instead of MSE(mean-squared error) term.

[royerloic]

Closing — Stale, but ideas documented for future work

Thank you @AhmetCanSolak — this PR introduces valuable CLI benchmark enhancements.

Why closing: The codebase has undergone significant changes since Dec 2022. The PR targets the old setup.cfg-era code and would not merge cleanly. Additionally, skimage.util.dtype.dtype_range may not be available in the current scikit-image 0.22+ version.

What's valuable and documented for re-implementation:

1. L1 loss function — Using MAE instead of MSE for self-supervised loss is a meaningful improvement:

   def loss_function(u, v):
       return numpy.mean(numpy.abs(u - v))

2. SSL PSNR metric — A novel metric combining PSNR with self-supervised loss:

   ssl_psnrs.append(
       10 * numpy.log10(dmax) - 10 * numpy.log10(ss_losses[-1])
   )

3. Pixel masking probability — Making pixel_masking_probability=0.1 explicit in the RandomMaskedDataset call.

These ideas are captured in code_review_reports/pr_ideas_for_future_work.md with full code context for re-implementation.