Fix EOS confidence suppression at nonzero temperature

[primorLee]

Summary

• suppress both EOS and EoT in the logits used to compute remasking confidence
• keep token sampling unchanged, so EOS/EoT can still be proposed normally
• add an order-sensitive regression test for the nonzero-temperature path

Problem

confidence_eos_eot_inf currently writes the EOS suppression to logits_with_noise after x0 has already been sampled. At nonzero temperature, logits_with_noise is a separate tensor, while low-confidence remasking computes probabilities from logits. EOS therefore retains its original confidence and can be transferred too early.

The zero-temperature path hides the defect because add_gumbel_noise returns logits itself, so the two names alias the same tensor.

Validation

• regression test fails on main: expected the ordinary token to transfer first, but EOS is fixed first
• python -m pytest tests/test_generate.py -q (1 passed)
• Python 3.12 compile check on Linux
• git diff --check

Fixes #137

[johnrunjiali-svg]

Thanks for the fix!