feat: Add histogram latency metrics for all inference phases

[zbennett10]

Summary

Adds 5 new Prometheus histogram metrics as equivalents to the existing 5 summary latency metrics. This gives operators bucket-based latency distributions that are more efficient for aggregation across instances in Prometheus compared to summaries.

New Metrics

Histogram	Corresponding Summary	Default Buckets (us)
nv_inference_request_duration_histogram_us	nv_inference_request_summary_us	1000, 5000, 25000, 50000, 100000
nv_inference_queue_duration_histogram_us	nv_inference_queue_summary_us	100, 1000, 5000, 10000, 50000
nv_inference_compute_input_duration_histogram_us	nv_inference_compute_input_summary_us	100, 500, 1000, 5000, 10000
nv_inference_compute_infer_duration_histogram_us	nv_inference_compute_infer_summary_us	1000, 5000, 25000, 50000, 100000
nv_inference_compute_output_duration_histogram_us	nv_inference_compute_output_summary_us	100, 500, 1000, 5000, 10000

Design

• Enabled via --metrics-config histogram_latencies=true (same flag as nv_inference_first_response_histogram_ms)
• Custom bucket boundaries configurable per-model via model_metrics.metric_control[] in config.pbtxt
• Request duration histogram is disabled when cache is enabled (matching the summary behavior, see DLIS-4762)
• Follows the exact same lifecycle as the existing TTFT histogram: registered in Metrics singleton, exposed via Family* accessors, initialized in MetricModelReporter::InitializeHistograms(), observed via ObserveHistogram()

Files Changed (6 files, +113/-4)

File	Change
src/constants.h	Add 5 new histogram constant names
src/metrics.h	Add 5 static family accessors + 5 member variables
src/metrics.cc	Register 5 new histogram families in constructor
src/metric_model_reporter.h	Add metric_map entries + default bucket boundaries
src/metric_model_reporter.cc	Wire histogram families in InitializeHistograms()
src/infer_stats.cc	Add ObserveHistogram() calls alongside existing summary observations

Motivation

Multiple users have requested histogram equivalents for the summary metrics (see #7672). Histograms are preferred over summaries for production monitoring because:

1. Aggregatable: Histogram buckets can be aggregated across instances; summary quantiles cannot
2. Predictable cost: Histograms have fixed memory overhead regardless of observation count
3. Alerting-friendly: histogram_quantile() in PromQL allows flexible percentile calculations from histograms

PR #8580 extended the TTFT histogram to non-decoupled models. This PR completes the histogram story by adding the remaining 5 latency metrics that users have been requesting.

Related Issues

• Addresses Histogram Metric for multi-instance tail latency aggregation server#7672
• Builds on PR #8580 (TTFT histogram for non-decoupled models)
• Companion PR: triton-inference-server/server (docs + test patterns)

Test Plan

• Pre-commit hooks pass (clang-format, codespell, trailing whitespace)
• Constructor initializer order matches member declaration order (no compiler warnings)
• Cache guard correctly applied to request_duration (matches summary behavior)
• All metric names consistent across registration, metric_map, and observation calls
• L0_metrics tests pass with Triton server (requires full build environment)
• Histogram output verified in /metrics endpoint with histogram_latencies=true

[zbennett10]

Companion PR -> triton-inference-server/server#8675

[whoisj]

@zbennett10, same as your other PR: please make sure you've completed the CLA requirements so that we can accept your contribution. Thanks!