Fix benchmark defects and flatten the benchmark package

[fpedd]

This fixes ten defects in the benchmark package and removes about 2000 lines from it: exported source labels that carry paths no longer split into nested directories, the tiling sources report their capacity as the known optimum instead of rebuilding the ground truth (and no longer return None without a seed), the subset sync pattern merges at least two threads, ONNX models take their id from the graph name and no longer inflate zero-size tensors, a campaign id of 0 is kept and the default save path no longer doubles its prefix, the Hugging Face variant hook no longer mutates the source mid-campaign, concatenated Hugging Face models get prefixed ids through a fixed-source base shared with Minimalloc, progress bars show the resolved allocator name, and PowerOf2Source checks its duration bound. run_benchmark now loops source, variant, allocator and records every skip in one list, the results package drops the double-built metadata and nested writers, Timer is a plain context manager, the Hugging Face tests build tiny local models instead of downloading 91 MB per test, and the examples are trimmed to minimal showcases of the API.