[skyrl-train] consider updating grpo_norm_by_std default to false

[erictang000]

Right now we set the default for trainer.algorithm.grpo_norm_by_std to true - for Dr. GRPO this should be false. Other research is adopting grpo_norm_by_std=false even if not using Dr. GRPO length normalization. We should consider updating this default.

Relevant discussion:

• Deepseek-v3.2 adopts this: https://x.com/zzlccc/status/1995770284385992798
• Qwen adopts this in MiniRL: https://arxiv.org/pdf/2512.01374

[erictang000]

Ablation (DAPO on Qwen3-30B-A3B, megatron tp=4, ep=8, with TIS + ratio 2) (light green is with grpo_norm_by_std=false)

• Generation length increases much more quickly
• Train reward is slightly higher, eval is slightly better but seems to be higher variance