Add data parallel support for vllm to speed up evals - #50

Given that evals take very long for large models, it would be nice to have data parallel working for vllm [see #31 #39].

My understanding is that the current codebase tries to address this with multi-processing, similarly to the vllm docs https://github.com/vllm-project/vllm/blob/main/examples/offline_inference/data_parallel.py.
However, this is currently broken #31 and additionally, unnecessary, since lm_eval already supports data parallel natively.

This PR should address these issues by adding lm_eval's data parallel support to Olmes. The first four commits are from #49 .

Tested with command on 4xh100s.

dp=2
tp=2
  olmes \
      --model "olmo-2-7b-1124" \
      --model-args '{"chat_model": true, "max_length": 4096, "metadata": {}, "trust_remote_code": true, "gpu_memory_utilization": 0.85, "data_parallel_size": '"$dp"', "tensor_parallel_size": '"$tp"'}' \
      --model-type vllm \
      --task-args '{ "generation_kwargs": { "max_gen_toks": 4096, "truncate_context": false } }' \
      --task gsm8k::tulu ifeval::tulu \
      --output-dir "eval_results"