no halting in evaluation

[yoyostudy]

TinyRecursiveModels/models/recursive_reasoning/trm.py

Line 275 in 7de0d20

if self.training and (self.config.halt_max_steps > 1):

trm model in training loop use q_head to introduce halting so it will not reach $N_{sup}=16$, however, in evaluation, there is no halting. It should match the convergent steps in training, otherwise the model will overthink.