stats: record Nighthawk statistics into an Envoy store histogram

[bpalermo]

Description

This PR is related to #1600

Records each Nighthawk statistic sample into an Envoy store histogram, so that Nighthawk's latency statistics appear in MetricSnapshot::histograms().

SinkableStatistic is constructed directly as a HistogramImplHelper over an empty MetricImpl and is never registered with the store, so it does not appear in a snapshot. Its only route to a sink was the direct Stats::Store::deliverHistogramToSinks() call in recordValue(). Sinks that implement onHistogramComplete() as an empty override and work from the snapshot on flush therefore received nothing for them — in Envoy that is metrics_service, open_telemetry and hystrix. Exporting a run through the OpenTelemetry sink would carry counters, gauges and Envoy's own histograms, with every Nighthawk latency metric absent and no error to indicate it.

ParentHistogramImpl::recordValue() calls deliverHistogramToSinks() itself, so recording into a store histogram replaces the direct delivery rather than adding to it: one call feeds both the snapshot and sinks that take individual samples.

Notes for Reviewers

• Emitted metric names are unchanged. The histogram is created in the worker's cluster.<n>. scope and named after the statistic's id, so sinks see cluster.<n>.benchmark_http_client.latency_2xx as before, and the worker remains recoverable from the name. A side effect is that Envoy's tag extractors now apply, so tag capable sinks receive the worker as envoy.cluster_name rather than only as a name prefix.
• The histogram is bound in setId(), since a statistic is named after construction, and the resolved pointer is cached so recording remains a dereference.
• Nighthawk's own output is unaffected: it is rendered from the HdrHistogram or Circllhist data and keeps nanosecond resolution. The mirror is a separate object and can carry a coarser representation later, once Envoy's statsd sinks scale by Histogram::Unit (stat_sinks: scale statsd histogram samples to milliseconds by unit envoy#47505), without changing what Nighthawk measures.
• Compatibility: NighthawkStatsSinkFactory and the onHistogramComplete() path are untouched, and test_stats_sinks.py passes unchanged. One exception worth naming: a sink recovering the worker id with dynamic_cast<const SinkableStatistic*> on the delivered histogram now receives a ParentHistogram and gets nullptr. The worker id remains available from the metric name or, for tag capable sinks, from the extracted tag.
• Cost: one mirrored record per request in HTTP mode, as the latency by status statistics are mutually exclusive. The added work is a cached thread local lookup and a circllhist insert, which measured below the run to run spread of the harness at 5k rps on loopback, against roughly 140 us of client CPU per request.
• The second commit adds //test:envoy_stats_sink_threading_test, which pins the threading contract Nighthawk must satisfy to host Envoy's own stats sinks: they resolve a thread local writer on every use, and every thread touching one must be registered with the ThreadLocal instance. Nighthawk's worker threads register in WorkerImpl's constructor and the main thread in ProcessImpl, including for a final flush issued after tls_.shutdownGlobalThreading(), which is the order ProcessImpl::shutdown() uses.
• Testing: //test:statistic_test asserts the statistic appears in store.histograms() of a ThreadLocalStoreImpl — note IsolatedStoreImpl::histograms() always returns an empty vector and cannot show this — plus //test:envoy_stats_sink_threading_test, //test:benchmark_http_client_test, //test:client_worker_test and //test:process_test.