Fix long task timer implementation

[tylerwowen]

The implementation of CumulativeHistogramLongTaskTimer in micrometer has a known issue. Besides, the implementation of writeLongTaskTimer is not compatible with CumulativeHistogramLongTaskTimer and results in non-sense histogram. The +Inf bucket should be the total counts of elements in the histogram, but it's set to the active count.

teletraan/deploy-service/universal/src/main/java/com/pinterest/teletraan/universal/metrics/micrometer/PinStatsPublisher.java

Lines 213 to 228 in cddbb72

	double count = timer.activeTasks();
	List<String> metrics = new ArrayList<>();
	metrics.add(writeMetricWithSuffix(timer.getId(), "active.count", wallTime, count));
	metrics.add(
	writeMetricWithSuffix(
	timer.getId(), "duration.sum", wallTime, timer.duration(baseTimeUnit)));
	metrics.add(writeMetricWithSuffix(timer.getId(), "max", wallTime, timer.max(baseTimeUnit)));
	if (percentileValues.length > 0) {
	metrics.addAll(writePercentiles(timer, wallTime, percentileValues));
	}
	if (histogramCounts.length > 0) {
	metrics.addAll(writeHistogram(wallTime, timer, histogramCounts, count, baseTimeUnit));

Because PinStats supports GaugeHistogram, this PR adopts the second recommendation in the micrometer issue.

Alternatively, GaugeHistogram is exactly what we want and doesn't require the logic of adding the previous snapshot to maintain monotonicity. Unfortunately, GaugeHistogram is not available in Prometheus' text format; it's only defined in OpenMetrics.

GaugeHistogram works for Teletraan's use case because the purpose of this metrics is to allow users to detect out of SLA host launches immediately. It's not intended to provide the histogram of actual host launch durations.

The fix is simply to inherit from DefaultLongTaskTimer instead of CumulativeHistogramLongTaskTimer.