RL + dLLM

[akoumpa]

BlockjustGRPO-fast + nemotron-lab-diffusion + AM + RL

AM integration Plan

https://docs.google.com/document/d/1xC7WKKNLlI5hEIly9--WelhO92NYKbgBMCqo62Po_1s/edit?usp=sharing

draft PR: zyzhou5/RL#4

**Plan **

1. Working e2e pipeline with AM backend (based on Sajad's current design https://github.com/sajadn/diffusion_RL/commits/dllm_clean/). Planning to do the following test
   1. model loading
   2. model forward/backward
   3. loss parity (algorithm)
   4. vLLM supports
   5. 100 steps of training and compare reward
2. Other features
   1. CP
3. Other algorithms
   1. JustGRPO
   2. …

Other things

1. Sajad mentioned about a new design of the dLLM RL pipeline in Nemo-RL (https://docs.google.com/document/d/1HHzRpLfsRalQkqLGFe0vljZjRO7dwKaKWH43jZPxCu4/edit?tab=t.0), we will need to refactor on the AM side as well once that new design is implemented in Nemo-RL
   1. Timeline - might happen in a few weeks
2. Check if huiying's refactor on AM + Nemo-RL will cause any issue refactor(engine): training Engine API #3614
   1. Timeline - estimated to be done by the end of September