serve_prm scoring service with vLLM module

[pranay5255]

Summary

Build serve_prm as the scoring service for branch prefixes. The interface should be usable from both offline eval and a future live forest reranker.

Scope

• Load a trained PRM checkpoint and expose batch prefix scoring.
• Support a clean API contract for branch_id, prefix, reward, rank_score, and optional rationale.
• Start with a reliable torch/transformers path; add vLLM-backed tokenization / generation integration where it makes sense, but do not force custom reward heads into a pure vLLM-only design if that makes the system brittle.
• Add a small client usable from eval scripts and future controller/runtime work.
• Add local smoke tests against saved prefixes from extracted traces.

Modules to build

• project/evmbench/evmbench/experiments/serve_prm.py
• project/evmbench/evmbench/experiments/prm_service.py
• project/evmbench/evmbench/experiments/prm_api.py
• project/evmbench/evmbench/experiments/prm_client.py
• project/evmbench/evmbench/experiments/vllm_backend.py

Acceptance criteria

• Given saved prefix inputs, the service returns deterministic scores with the expected schema.
• Batch scoring works for multiple live branches.
• There is a documented backend story for custom reward heads vs vLLM generation.