benchmarking: expand cAdvisor scrape allowlist for kernel PSI, networ…

[Nishanth Kotla (Nishanth29)]

Fixes #1588

What this PR does

This PR updates the Prometheus cAdvisor scrape configuration in benchmarking/monitoring.yaml to retain Linux kernel Pressure Stall Information (PSI), network bandwidth, disk I/O throughput, and host-level node resource utilization.

Problem

Currently, monitoring.yaml restricts cAdvisor scrapes to only:
container_(memory_working_set_bytes|cpu_usage_seconds_total|spec_memory_limit_bytes)
and unconditionally drops series with container="".

Because cAdvisor reports the host root cgroup (id="/") with an empty container label, this configuration strips:

1. All host-level kernel PSI stalls (cpu.pressure, io.pressure, memory.pressure).
2. Host system daemon resource overhead (containerd, kubelet, virtualization runtime).
3. Container network bandwidth (container_network_*) and disk I/O (container_fs_*).

Under high-density multi-tenant agent benchmarks, diagnosing whether tail latency spikes stem from CPU scheduling contention, disk storage bottlenecks, or memory pressure is difficult without kernel PSI.

Proposed Changes

In benchmarking/monitoring.yaml:

1. Relabel Host Root cgroup: Relabel id="/" to container="node" before dropping empty container labels, preserving host-level node rollups.
2. Expand Allowlist Regex: Keep container_pressure_* (Linux PSI), container_network_*, container_fs_*, container_cpu_cfs_*, and container_memory_rss.
3. Prevent Pod Double-Counting: Drop empty container labels selectively only for cumulative memory working set and CPU usage to prevent 2x double-counting between pod slices and container scopes.

How this was tested

Tested on a benchmark cluster running the glutton user workload:

• Verified that container_pressure_* metrics (CPU, memory, and I/O) are actively collected and queryable in Prometheus.

• Verified that host-level root cgroup metrics appear under container="node".

• Verified that container network and filesystem I/O metrics are retained, while pod-level duplicate CPU and memory working-set series are dropped as intended.

• Tests pass

• Appropriate changes to documentation are included in the PR

[Max Smythe (maxsmythe)]

LGTM