scx_p2dq: Add L2 cluster-aware CPU selection

[hodgesds]

Add L2 cluster awareness to improve cache locality by preferring CPUs
within the same cluster before searching the wider LLC domain.

Infrastructure:

• Add cluster_id to cpu_ctx for per-CPU cluster tracking
• Add has_clusters flag to topo_config
• Initialize cluster_id for each CPU during BPF setup
• Populate cluster IDs from topology in userspace

Implementation:

• Add pick_idle_cpu_in_cluster() helper to search cluster cpumask
• Enhance pick_idle_cpu() to try cluster-level before LLC-level
• Update wakeup paths for interactive tasks to prefer cluster
• Check same-cluster waker/wakee before wider search

This improves cache locality by keeping related tasks on CPUs sharing
L2 cache, reducing cache misses and improving performance.

[hodgesds]

Comparison: Optimized Cluster-Aware vs Baseline

schbench Latency Test (2 threads × 16 messages, 30s) - 5 Run Average

Metric	Baseline	Optimized Cluster-Aware	Change
Average RPS	7,801.15	7,784.21	-0.22% ✓
RPS Range	7,771 - 7,830 (59 RPS)	7,727 - 7,836 (109 RPS)	+1.8x variance
Std Deviation	~25 RPS (0.32%)	~48 RPS (0.61%)	+1.9x
Wakeup Latency p99	15 µs	14 µs	-6.7% ✓
Request Latency p99	7,720 µs	7,704 µs	-0.2% ✓

Result: Near-identical performance with baseline in low-contention scenarios.

---

stress-ng Cache Benchmark (60s, 316 workers)

Metric	Baseline	Optimized Cluster-Aware	Change
Cache ops/sec	36,241,353	38,130,668	+5.21% ✓
Cache writes/sec	2,777,973	3,931,893	+41.5% ✓✓

Result: Significant cache performance improvement from L2 locality.

---

stress-ng CPU Benchmark (60s, 316 workers)

Metric	Baseline	Optimized Cluster-Aware	Change
Bogo ops/sec	332,453	333,337	+0.27% ✓

Result: No CPU throughput regression.

---

schbench High Load Test (8 threads × 32 messages, 30s)

Metric	Baseline	Optimized Cluster-Aware	Change
Average RPS	31,773	31,923	+0.47% ✓
Wakeup Latency p99	22 µs	21 µs	-4.5% ✓
Request Latency p99	9,648 µs	9,584 µs	-0.66% ✓

Result: Improved performance under high contention.

---

Key Findings

✅ Wins

1. Cache Performance: +5.21% overall, +41.5% cache writes
2. High-Load Performance: +0.47% RPS, better latencies
3. Low-Load Performance: -0.22% RPS (essentially baseline)
4. Wakeup Latency: 4.5-6.7% faster across all scenarios
5. CPU Throughput: Neutral (+0.27%)

⚠️ Minor Considerations

1. Variance: Slightly higher variance in low-load scenarios (1.9x vs baseline)
   - Still very acceptable at 0.61% standard deviation
   - Baseline has exceptionally low variance (0.32%)

[hodgesds]

clangd-format messed up some of the bpf code, will push a fix.