feat(hash): add compare command, bundle inputs, and plot --kind/--scale

[helly25]

Summary

Measurement-tooling additions to hash_benchmark_report.py, a denser latency benchmark, and reporting/consistency polish.

New commands & inputs

• compare — per-case Δ% (with a geomean summary row) between two datasets; negative = B faster (lower ns). Shows both machine labels; only cases present in both datasets are compared.
• Unified inputs — tables/plot/compare/quality accept a bundle .tgz or a canonical results JSON interchangeably, positionally or via --results/--bundle (auto-detected by extension). Existing flag invocations still work.

plot

• --kind {throughput,latency,all} — latency is now dense enough to render as a log-log curve.
• --scale {log-log,linear-log} — linear-log = linear y / log x (nice 0-based ticks) to read absolute ns gaps; x stays log always.

Benchmark (hash_benchmark.cc)

• Expanded kFullSizes (with rationale comments on the boundaries) and made the latency benchmark sweep the full size set, so latency plots as a dense ns-vs-length curve.

Reporting polish

• publish orders sections and the manifest by the generated header (machine label), so the README is stable regardless of the order bundles are passed.
• Latency table uses the curated readme_sizes subset, like the throughput tables (no longer dumps every length now that latency sweeps the full set).
• Size labels unified everywhere (tables + chart x-axis): consistent B/KiB suffix (e.g. 1 KiB, 4 KiB); the latency column now uses the same labels instead of raw numbers.
• Machine label / headers now include the OS (MacOS/Linux).
• Throughput and latency table columns are right-aligned.

Fixes

• Escaped % as %% in the compare help string — a literal % is a printf format for argparse, which Python 3.14 validates eagerly at parser-construction time, so the tool crashed on 3.14 before this.

Compatibility / follow-up

• plot log-log output is byte-identical to before (the tool's own verify passed on the prior chart set).
• The OS/label and size-label changes alter chart subtitles and table text, so the committed README perf section + charts need a publish regen (done separately with the final bundle set); verify covers it.