docs: stale performance claims in BENCHMARK.md and USER_GUIDE.md

[heynemann]

Stale performance claims remaining in docs

While fixing the README/PERFORMANCE.md claims in #8, I found that BENCHMARK.md and USER_GUIDE.md carry the same class of stale numbers. These were left untouched to keep that PR reviewable; this issue tracks the follow-up.

BENCHMARK.md

• Claimed throughput figures (e.g. "84.7M msg/s", "77.1M msg/s") predate the throughput-benchmark timing fix and were measured with a fixed 500ms sleep inside the timed window. True end-to-end fan-out measured after the fix: 9.9–10.7M deliveries/s (direct goroutines), 12.4–12.8M/s (worker pool), 1000 subscribers, Apple M4 Pro, Go 1.26.6.
• "30–50x faster than MQTT" and "1,000–5,000x faster" wildcard claims are not reproducible. Measured vs MochiMQTT in-process routing: ~1.04x faster single publish, ~1.9x faster concurrent publish, 7–11x less memory per publish, 3.5x slower subscribe/unsubscribe churn.

USER_GUIDE.md

• Repeats the "84.7M msg/s" and "10–13% faster" delivery-mode figures; the measured crossover is the reverse at 1000 subscribers (worker pool slightly faster than direct goroutines).
• MaxConcurrentSubscriptions guidance ("optimal 5–300 worker pool / 1–750 direct", "up to 34x worse when set too high") is unverified in this session.

Suggested fix

Re-run the affected benchmarks and rewrite these files to cite the same measured ranges used in README/PERFORMANCE.md, or trim them down to qualitative guidance with a pointer to PERFORMANCE.md.