[RL] Fix rl on main

[Lucaskabela]

Following the config changes in #2386 the scripts for RL fail on main.

This PR modifies the code to restore them

Test Plan

Simple RL

VLLM_BATCH_INVARIANT=1 VLLM_ATTENTION_BACKEND=FLASH_ATTN python3 torchtitan/experiments/rl/vllm_compat/simple_rl.py

output:

Step   0 | Loss: -0.0036 | Reward: +0.075 | Samples: 160
  Sample:  48 + 24 = 72.
Let me check the reasoning.
Okay, so Natalia sold clips to 48 of ...
...

Simple Multiprocess RL

VLLM_BATCH_INVARIANT=1 VLLM_ATTENTION_BACKEND=FLASH_ATTN python3 torchtitan/experiments/rl/unified/simple_rl_multiprocess.py

[2026-02-23 15:03:25] INFO trainer.py:114: [actor=<root>.<torchtitan.experiments.rl.unified.actors.trainer.Trainer trainer{'gpus': 0/2}>] os.getpid()=3548191 Trainer starts to train 1 on traj:
  ✓ vLLM-TorchTitan bitwise determinism verified: 20 tokens match exactly
  ✓ vLLM-TorchTitan bitwise determinism verified: 20 tokens match exactly

[wwwjn]

I updated this PR to use new Config System, you can use it as a reference for the RoPE part : #2191

But this PR contains more - It redesigns the RL side config system so it might go through several rounds of review. If you want to land your fix first, I'm also ok with it