fix: treat temperature=0 as greedy selection, untempered logprobs

[EazyReal]

Problem

get_responses and get_log_probs_and_entropy divide logits by rollout_temperature whenever it is not 1.0. Greedy rollout uses temperature=0, so that is a divide-by-zero.

SGLang treats T=0 as greedy token selection, but still returns finite untempered model log-probs. slime stores those as rollout_log_probs and builds PPO ratios from exp(log π_θ − log π_old). Forcing greedy log-probs to 0 would make an on-policy sample start at r = π_θ(a).

Value heads already pass apply_temperature=False.

Change

apply_rollout_temperature is the shared rule at both logit sites:

• T in {0, 1}: leave logits unchanged
• otherwise: logits / T

T=0 also skips top-p nucleus replay. SGLang remaps T=0 to top_k=1 and reports unmasked untempered log-probs.

Test plan

• tests/test_rollout_temperature.py — recomputed logprobs match log_softmax at T=0/1 and log_softmax(logits/T) at T=0.5; get_responses does not scale at 0; greedy skips top-p replay
• tests/test_value_temperature.py — value head still ignores rollout temperature

[EazyReal]

reviewer @zhuzilin commented offline that rollout_temperature must be > 0 in RL to ensure meaningful exploration, so we do not need to consider this case.
opened and merged #2297 to ensure that.