Optimize exporter CPU and memory usage for high-density nodes

[jzwlqx]

Summary

This MR addresses high CPU (~367m) and memory (~300Mi) consumption of kubeskoop-agent on nodes
with many pods (60+). Through a series of targeted optimizations, CPU is reduced by 72% to ~101m
and memory by 58% to ~126Mi.

Changes

1. LinkManager abstraction (linkmanager.go)

• Centralized network interface management with netlink subscription for link events
• Replaced per-probe ad-hoc interface discovery in flow and nlqdisc probes
• Eliminated redundant netlink connections and polling

2. Switch from bswang/netlink fork to upstream vishvananda/netlink

• The forked netlink library had a busy-loop bug (nonblocking socket without epoll registration)
  causing unnecessary CPU consumption
• Migrated to the official upstream library

3. Optimize proctcpsummary /proc/net/tcp parsing (proctcp.go)

• Replaced regexp-based parsing with single-pass zero-allocation field extraction
• Added 2-second TTL cache to avoid redundant collection across metrics

4. NETLINK_SOCK_DIAG for TCP socket collection (sockdiag.go)

• Implemented binary SOCK_DIAG protocol (inet_diag) as replacement for text-based /proc/net/tcp
• 2.3x faster per-collection (713ms -> 306ms avg), 2.34x less kernel CPU
• Automatic fallback to /proc parsing if SOCK_DIAG is unavailable
• Benchmark at scale: up to 5x speedup at 10K+ sockets (p50)

5. Lazy-load kernel symbols (kallsyms.go)

• Changed unconditional init() loading of 214K symbols (~30MB) to sync.Once lazy loading
• Symbols only loaded when BPF probes (tracepacketloss, tracetcpretrans) are active
• Saves ~30MB heap for the common case where no BPF probes are enabled

6. Eliminate intermediate string map allocations (procnetstat.go, procsnmp.go)

• procnetstat: replaced map[string]map[string]string -> direct uint64 parsing, only extracting
  needed fields. Reduced cumulative allocations from ~7.7GB to near zero.
• procsnmp: same optimization for /proc/net/snmp (tcp/udp/ip protocols). Reduced ~550MB
  cumulative allocations.

Performance Results (node with 63 pods, 59 unique namespaces)

Metric	Before	After	Reduction
Pod CPU	367m	101m	-72%
Pod Memory (RSS)	~300Mi	126Mi	-58%
Go Heap (live)	52.7MB	32MB	-39%
GC cycles (2min)	1431	304	-79%
TCP collection (p50)	687ms	210ms	-70%
Kernel CPU per coll.	38.3s	16.4s	-57%

Test Plan

• Unit tests for SOCK_DIAG parsing (sockdiag_test.go) and proc parsing (proctcp_test.go)
• All 10 configured probes verified producing correct metrics post-deploy
  (conntrack, qdisc, netdev, io, sock, tcpsummary, tcp, tcpext, udp, rdma)
• Metrics cross-validated between problem node (.51, 63 pods) and normal nodes
• pprof heap/cpu profiling before and after each optimization
• Standalone benchmark tool (cmd/tcpbench) validated SOCK_DIAG vs proc at
  0/1K/10K/50K socket scales
• No new errors in inspector logs (rdma errors are pre-existing, unrelated)