feat(nccl): add nccl-tests all_reduce_perf for GPU bandwidth testing

- Dockerfile: install cuda-nvcc-13-0 from NVIDIA repo for compilation
- build-nccl-tests.sh: downloads libnccl-dev for nccl.h, builds all_reduce_perf
- build.sh: runs nccl-tests build, injects binary into /usr/local/bin/
- platform: RunNCCLTests() auto-detects GPU count, runs all_reduce_perf
- TUI: NCCL bandwidth test entry in Burn-in Tests screen [N] hotkey

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

audit/internal/platform/sat.go

@@ -121,6 +121,24 @@ func (s *System) ListNvidiaGPUs() ([]NvidiaGPU, error) {
 	return gpus, nil
 }
 
+// RunNCCLTests runs nccl-tests all_reduce_perf across all NVIDIA GPUs.
+// Measures collective communication bandwidth over NVLink/PCIe.
+func (s *System) RunNCCLTests(ctx context.Context, baseDir string) (string, error) {
+	// detect GPU count
+	out, _ := exec.Command("nvidia-smi", "--query-gpu=index", "--format=csv,noheader").Output()
+	gpuCount := len(strings.Split(strings.TrimSpace(string(out)), "\n"))
+	if gpuCount < 1 {
+		gpuCount = 1
+	}
+	return runAcceptancePackCtx(ctx, baseDir, "nccl-tests", []satJob{
+		{name: "01-nvidia-smi-q.log", cmd: []string{"nvidia-smi", "-q"}},
+		{name: "02-all-reduce-perf.log", cmd: []string{
+			"all_reduce_perf", "-b", "512M", "-e", "4G", "-f", "2",
+			"-g", strconv.Itoa(gpuCount), "--iters", "20",
+		}},
+	})
+}
+
 func (s *System) RunNvidiaAcceptancePack(baseDir string) (string, error) {
 	return runAcceptancePack(baseDir, "gpu-nvidia", nvidiaSATJobs())
 }