Add configurable temperature parameter for RL rollout sampling

[Xiuyu-Li]

This PR adds temperature as a configurable argument in Config, allowing users to adjust the sampling temperature used for the RL training. It also updates the math_rl example (tinker_cookbook/recipes/math_rl/train.py) to demonstrate how to easily apply this setting within a training script.

Note: This change only affects the TinkerTokenCompleter class, which is the default completer used by most RL recipe scripts. The TinkerMessageCompleter class (in tinker_cookbook/completers.py) remains unchanged — it still uses a hard-coded temperature=1.0 in its SamplingParams, left as-is for now to avoid altering potential design-specific behavior. As a result, this update does not impact scripts that rely on TinkerMessageCompleter, such as text_arena, twenty_questions, and similar examples.

[Tiiiger]

LGTM! I'll click on merge once I can confirm the backend actually supports this correctly.

[joschu]

I'm not sure logprobs are calculated the right way when temperature != 1 -- for this to make sense, we'd need for the sampling engine to compute the logprob of the temperature-scaled model, not the original model.
Have you done experiments @Xiuyu-Li?
Marking as "Requesting Changes" until we've validated that this makes sense.