Scan time budget and performance monitor add CPU on every pattern check #62

Every pattern check pays for the scan time budget: checkRegexPatterns reads ThreadMXBean.currentThreadCpuTime twice per pattern check, and so does the fallback scan. The performance monitor also copies the recent-times window twice and allocates a list for the variance on every record, and sanitizes each anomaly before it knows whether any callback is registered. The Python reference engine fixed the same overhead in Guard-Core/guard-core#172. I am charging the budget per pass at 2 ms slice boundaries and making the monitor statistics allocation-free.