pyperf JSON hoists identical per-process metadata to the benchmark level

When reading a pyperf/pyperformance JSON, per-worker metadata such as runnable_threads, load_avg_1min or even date is only stored in runs[i].metadata if it differs between processes. If every process saw the same value, pyperf moves the key into benchmarks[j].metadata (and values shared by all benchmarks can sit in the top-level metadata). Code that only does run['metadata'].get('runnable_threads') silently gets None for exactly the quiet benchmarks. Resolve keys with run metadata, then benchmark metadata, then file metadata. Seen in faster-cpython's public pyperformance results, where runnable_threads appears only at the benchmark level.