Fix LFW eval distance shape

[ZardLi1115]

Summary

• Add a shared LFW distance helper that flattens each model output to one feature vector per sample before computing pair distance.
• Use the helper in both standalone LFW evaluation and training-time LFW validation so distance arrays stay one-dimensional and match the label array.

Tests

• python -m py_compile utils/utils_metrics.py utils/utils_fit.py
• python inline shape check for calculate_lfw_distance with fake 4D outputs (workspace lacks numpy/torch)
• git -c core.whitespace=blank-at-eol,blank-at-eof,space-before-tab,cr-at-eol diff --check

Fixes #46