Fix eval concurrency argument compatibility

[haofrank]

Summary

Fix accuracy evaluation failures caused by passing the unsupported
--concurrent-requests argument to InferenceX run_eval, while ensuring
accuracy artifacts are reliably persisted for both local and remote evaluation.

Changes

• Configure lm-eval concurrency through EVAL_CONCURRENT_REQUESTS.
• Add MAGPIE_EVAL_CONCURRENCY to control accuracy concurrency independently
  from performance benchmark concurrency.
• Apply the concurrency fix consistently to SGLang, vLLM, and ATOM benchmark
  scripts on MI300X and MI355X.
• Use the following concurrency precedence:
  1. MAGPIE_EVAL_CONCURRENCY
  2. EVAL_CONCURRENT_REQUESTS
  3. benchmark CONC
• Preserve raw lm-eval JSON/JSONL artifacts under lm_eval/.
• Generate a stable accuracy_report.json beside benchmark_report.json.
• Support batched concurrency values such as
  EVAL_CONCURRENT_REQUESTS="2 4" without losing staged results.
• Record the actual accuracy concurrency in meta_env.json, rather than the
  performance benchmark concurrency.
• Propagate accuracy-report persistence failures for remote evaluations instead
  of reporting a false success.
• Leave source lm-eval artifacts untouched when copying them into the benchmark
  workspace.
• Update benchmarking documentation and the Magpie skill reference.

Output

When accuracy evaluation is enabled, the benchmark workspace now includes:

results/benchmark_<framework>_<timestamp>/
├── benchmark_report.json
├── accuracy_report.json
└── lm_eval/
    ├── results_*.json
    ├── samples_*.jsonl
    └── meta_env.json

[sinarafati-amd]

looks good

[sinarafati-amd]

looks good