Currently we don't support batch_size<dp_size and batch_size % shards != 0. We should relax this so that users do not have to recreate the vllm engine or the policy if they have an odd batch_size.

This leads to a UX where we have to tear down the worker groups each time we want a different batch size which isn't necessary.

Examples:

• vllm: https://github.com/NVIDIA/reinforcer/blob/d9277a8afcc2f4892ebe7225754b94b586de09d9/nemo_reinforcer/models/generation/vllm.py#L542
• hf: https://github.com/NVIDIA/reinforcer/blob/d9277a8afcc2f4892ebe7225754b94b586de09d9/nemo_reinforcer/models/policy/hf_policy.py#L934