Scale RL features

[terrykong]

• advantage normalization (batch level)
• zero-variance filtering
• FP32 lm head (train and gen)
• CISPO
• no-positive resampling (data curriculum: skip >=0.9 pass rate)

cc/ @parthchadha