macOS: benchmark run can hang indefinitely with default multiprocessing/data loader settings

Description

On macOS, benchmark_runner may hang indefinitely during benchmark run when using default runtime settings.
The process reaches the benchmark phase but does not make progress for several minutes (or longer), and no explicit error is raised.

Environment

• OS: macOS (Apple Silicon)
• Python: 3.11
• benchmark-runner: v0.0.2
  Backend: OpenAI-compatible HTTP endpoint
  Data: synthetic (prompt_tokens=...,output_tokens=...)

Reproduction

python -m benchmark_runner.main benchmark run \
  --target http://<server>:<port> \
  --sample-requests 0 \
  --output-dir .cache \
  --outputs full.json \
  --processor <local_tokenizer_path> \
  --data prompt_tokens=8192,output_tokens=100 \
  --profile constant \
  --rate 1 \
  --max-requests 10

Expected behavior

Benchmark starts and either: progresses normally, or fails fast with a clear error.

Actual behavior

The command appears stuck/hung (often at the benchmark stage) without completion or explicit failure.

Start with the python -m benchmark_runner.main benchmark run entry point and reproduce the hang on macOS with the command shown, focusing on the default multiprocessing and data-loader settings. Trace the benchmark phase to determine why it neither progresses nor fails; done means the run completes normally or exits with a clear error under those defaults.