Per-loss logits_scale_factor

[jlamypoirier]

Summary

Add a logits_scale_factor field to LanguageModelLossConfig, applied on top of the head's logits_scale_factor for that loss only. Every loss subclass picks it up automatically via self._logits_scale_factor — no per-subclass changes.

Primary use case: in RL losses, set to 1 / actor_temperature so new log-probabilities are computed at the same scale as the actor's stored old log-probabilities (importance ratio at step 0 is no longer offset by the actor's sampling temperature).

Reimplemented from PR #502's temperature field on the GRPO config: the field is moved to the base config (so all losses can opt in), renamed to match the existing head field, and given multiplier semantics (extra scale, default 1.0, stacked on top of the head's scale).

Test plan

• Existing loss tests still pass at default logits_scale_factor=1.0.
• Manual: setting logits_scale_factor=2.0 on a GRPO loss config produces softmax outputs equivalent to doubling the head's logits_scale_factor.