Evaluation results differ when `num_prompts_per_step` is different

**Describe the bug**

Evaluation results differ when `num_prompts_per_step` is different.

**Steps/Code to reproduce bug**

In the `main` branch, run the two following commands and the results are different:

- `uv run examples/run_eval.py data.dataset_name="HuggingFaceH4/MATH-500" data.dataset_key=test`. This will a score of 76.2. The default value for `num_prompts_per_step` is -1.
- `uv run examples/run_eval.py data.dataset_name="HuggingFaceH4/MATH-500" data.dataset_key=test generation.num_prompts_per_step=128`. This will give a score of 75.6.

Both used greedy decoding.

**Expected behavior**

See above.

**Environment overview (please complete the following information)**

- Environment location: Docker
- Method of install: Building the Release docker container (https://github.com/NVIDIA-NeMo/RL/blob/main/docker/Dockerfile)
- If method of install is [Docker], provide `docker pull` & `docker run` commands used: N/A

**Environment details**

If NVIDIA docker image is used you don't need to specify these.
Otherwise, please provide:
- OS version
- PyTorch version
- Python version

**Additional context**
N/A