Support multi-node evaluation with vLLM

Problem description:

• Currently evalchemy supports only multi-gpu (and not multi-node): data-parallel, based on accelerate and tensor-parallel, based on HF transformers evaluation. These 2 options are mutually exclusive. There's already a PR that expands on the accelerate approach to support multi-node setup. This approach seems to work fine with small models.
• However, for larger models the approach that uses accelerate is not feasible.

Proposed solution:

• Support vLLM backend (based on Ray) like lm-eval-harness.
• vLLM supports highly optimized multi-node inference and both data- and tensor-parallelism.
• In principle, vLLM should give 2 advantages: 1) faster evaluations than HF, especially with batched inference and 2) support for bigger models that do not fit on one GPU.

Start by reviewing PR #29 and the linked lm-eval-harness vLLM integration, then read the vLLM and Ray documentation linked in the issue. The work is complete when evalchemy has a vLLM backend that supports multi-node inference with data- and tensor-parallelism for larger models.