fix: avoid applying rollout temperature to critic values

[Baiyu-Su]

What

Avoid applying rollout_temperature to critic value-head outputs.

get_responses() is shared by response-aligned policy-logit extraction and value extraction. Temperature scaling is needed for policy logits when reconstructing rollout log probabilities, but critic values are scalar predictions and should not be divided by the rollout sampling temperature.

Changes

• Add an apply_temperature flag to get_responses().
• Keep temperature scaling enabled by default for existing policy/logprob paths.
• Disable temperature scaling in get_values().
• Add a zero-GPU unit test for non-unit rollout_temperature.

Tested

• ruff check slime/backends/megatron_utils/loss.py tests/test_value_temperature.py
• isort --profile=black --filter-files --check-only slime/backends/megatron_utils/loss.py tests/test_value_temperature.py
• black --check slime/backends/megatron_utils/loss.py tests/test_value_temperature.py
• PYTHONDONTWRITEBYTECODE=1 python -m py_compile slime/backends/megatron_utils/loss.py tests/test_value_temperature.py