Report cumulative vLLM generation progress across batches

[finbarrtimbers]

Summary

• Add BeakerStatusReporter.start_accumulator / accumulating_callback so progress can be reported as a single cumulative counter across batches.
• Wire VLLMServerProvider to use the accumulator when the runner exposes OLMO_VLLM_GEN_TOTAL, falling back to per-batch reporting otherwise.
• Runner exports OLMO_VLLM_GEN_TOTAL (sum of total_instances * num_samples) and OLMO_EVAL_BENCHMARK so the provider can label progress with the experiment name.

Test plan

• Small vLLM run; confirm the progress bar reflects cumulative generation count instead of resetting per batch.

🤖 Generated with Claude Code