fix: normalize category labels in functional nominal metrics for non-zero-based labels

[xiaoyaoqilan]

Problem

Functional nominal metrics (cramers_v, pearsons_contingency_coefficient, theils_u, tschuprows_t) crash when category labels are not zero-based and contiguous:

cramers_v(torch.tensor([1, 1, 2, 2]), torch.tensor([1, 1, 2, 2]), bias_correction=False)
# RuntimeError: shape '[2, 2]' is invalid for input of size 7

Nominal association is invariant to renaming categories — labels [0,1] and [1,2] describe the same two categories and should produce the same result.

Root Cause

num_classes is inferred as the count of unique values (2), but the confusion-matrix updater uses raw label values as bin indices. Label 2 creates a bin outside the expected 2x2 range.

Fix

Added _normalize_categorical_labels to nominal/utils.py which maps observed labels to contiguous 0-based IDs when they aren't already. Called from all 4 update functions after NaN handling:

• cramers.py
• pearson.py
• theils_u.py
• tschuprows.py

Stateful metrics with explicit num_classes are unaffected (they already use stable encoding).

Verification

base = torch.tensor([0, 0, 1, 1])
shifted = base + 1
# Before: shifted input raises RuntimeError
# After: both produce tensor(1.) for Cramer's V

Fixes #3446