docs(irodori-tts): add v4.1-Small models

v4.1-Small differs from v4-Small only in the duration predictor, which
upstream retrained separately with every other parameter frozen: 683 of
the 714 tensors are bit-identical and the config is unchanged.

Refresh the caption-only duration table with v4/v4.1 measurements taken
under matched conditions, and record the PyTorch cross-check (117.41
frames against MLX's 117.56, both rounding to 118) so the remaining
over-prediction is documented as upstream behaviour rather than a port
artifact.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>