Inference strategy bug for looplm

[ACautomata]

LoopLM/src/inference/early_exit.py

Lines 194 to 195 in 8d79e31

	# CDF update (scalar — mean over batch and positions)
	lam_mean = lam.mean().item()

It seems that all samples in a batch will exit at the same step. The official implementation of loop exit is token-level could be found here. https://huggingface.co/ByteDance/Ouro-1.4B/blob/main/modeling_ouro.py#L781