feat(simclr): 接入 SwanLab 可视化，并补上本地看板缺失的依赖

- SimCLRTrainer 记录 simclr/loss、loss_avg、lr（按 step）与 epoch_loss、
  epoch_seconds、gpu_mem_peak_mb（按 epoch），超参为完整 yaml 配置；
  正常结束与 KeyboardInterrupt 两条路径都会 finish()
- 配置新增 simclr.swanlab（字段同 train.swanlab），默认 mode: local，
  不需要登录，日志写到 <work_dir>/swanlog
- 依赖改为 swanlab[dashboard]：此前 `swanlab watch` 在干净环境里会报
  "Please install the swanboard package"，README 里检测训练那节的看板命令
  同样受影响，现已随 uv sync 装好
- README 补齐本地看板/云端/离线补传三种用法与端口、局域网参数

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

engine/simclr.py

@@ -18,6 +18,7 @@ from torch.utils.data import DataLoader
 
 from engine.metrics import AverageMeter
 from engine.optim import build_optimizer, build_scheduler
+from engine.visualizer import SwanLabLogger
 from models.heads.projection import ProjectionHead
 from utils.device import PeakMemory, resolve_amp, synchronize
 from utils.logging import get_logger
@@ -45,7 +46,9 @@ class SimCLRTrainer:
     """Args:
     backbone: 输出特征图的卷积骨架（需带 `out_channels`）。
     dataloader: 每个样本产出 2 个视图的 DataLoader（`UnlabeledImageDataset(views=2)`）。
-    cfg: `epochs / temperature / optimizer / scheduler / amp / log_interval / save_interval` 等。
+    cfg: `epochs / temperature / optimizer / scheduler / amp / log_interval / save_interval`
+        以及 `swanlab`（可视化，字段同 `train.swanlab`）。
+    full_config: 记录到 SwanLab 的超参，一般直接传完整 yaml 配置。
     """
 
     def __init__(
@@ -55,6 +58,7 @@ class SimCLRTrainer:
         cfg: Mapping[str, Any],
         device: torch.device,
         work_dir: str | Path,
+        full_config: dict | None = None,
     ) -> None:
         self.backbone = backbone
         self.dataloader = dataloader
@@ -81,6 +85,13 @@ class SimCLRTrainer:
         self.scaler = torch.amp.GradScaler(device.type, enabled=self.amp)
         self.peak_memory = PeakMemory(device)
 
+        self.global_step = 0
+        self.visualizer = SwanLabLogger(
+            work_dir=self.work_dir,
+            config=full_config,
+            **dict(self.cfg.get("swanlab", {})),
+        )
+
     def train_one_epoch(self, epoch: int) -> float:
         self.model.train()
         meter = AverageMeter()
@@ -101,7 +112,17 @@ class SimCLRTrainer:
 
             self.peak_memory.update()
             meter.update(float(loss.detach()), view1.shape[0])
+            self.global_step += 1
             if step % self.log_interval == 0:
+                self.visualizer.log(
+                    {
+                        "loss": float(loss.detach()),
+                        "loss_avg": meter.avg,
+                        "lr": self.optimizer.param_groups[0]["lr"],
+                    },
+                    step=self.global_step,
+                    prefix="simclr/",
+                )
                 logger.info(
                     "simclr epoch %d [%d/%d] loss=%.4f lr=%.2e",
                     epoch,
@@ -113,11 +134,22 @@ class SimCLRTrainer:
 
         synchronize(self.device)
         peak = self.peak_memory.peak_mb
+        elapsed = time.time() - start
+        self.visualizer.log(
+            {
+                "epoch": float(epoch),
+                "epoch_loss": meter.avg,
+                "epoch_seconds": elapsed,
+                "gpu_mem_peak_mb": peak,
+            },
+            step=self.global_step,
+            prefix="simclr/",
+        )
         logger.info(
             "simclr epoch %d 完成 loss=%.4f 耗时 %.1fs 显存峰值 %.0f MB%s",
             epoch,
             meter.avg,
-            time.time() - start,
+            elapsed,
             peak,
             "" if self.peak_memory.exact else "（MPS 采样下界）",
         )
@@ -149,8 +181,10 @@ class SimCLRTrainer:
         except KeyboardInterrupt:
             self.save_backbone(epoch)
             logger.warning("训练被中断，已保存 epoch %d 的权重: %s", epoch, path)
+            self.visualizer.finish()
             return path
 
         self.save_backbone(self.epochs - 1)
         logger.info("自监督骨架已保存: %s", path)
+        self.visualizer.finish()
         return path