GRPO OOM testing script

For testing ooms, right now our current process is to just launch a regular GRPO run and wait for it to oom after a few minutes. this process may take 5-10 mins or more after the job is running depending on the config. it may be helpful to make a grpo oom-testing script which skips the initial vllm cuda graph capture, validation, and rollouts and just runs through the policy train with the configured global/micro batch sizes and seq lens.

A more concrete definition: given an input config, do the regular GRPO setup but do not generate rollouts. In place of the rollouts, create dummy tensors using the max seq len and try to take one train step.