fix(simclr): 换优化器种类时不再加载不兼容的动量缓冲

adamw 换成 sgd 时 load_state_dict 不会报错，SGD 会收下 AdamW 的 exp_avg/exp_avg_sq
却永远用不上——动量实际从零开始，但状态里白占内存，之后存盘还会一路带下去。
改为检测到优化器种类变化就跳过加载并给出日志，模型权重与训练进度不受影响。

README 补一张表，说明续训时哪些配置按 yaml、哪些来自存档、哪些不能改。

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

tests/test_simclr.py

@@ -330,6 +330,33 @@ def test_resume_takes_hyperparams_from_config_not_checkpoint(tiny_image_dir, tmp
     assert resumed.optimizer.state, "动量缓冲应当保留——那才是真正要续的'进度'"
 
 
+def test_resume_across_optimizer_types_drops_incompatible_state(tiny_image_dir, tmp_path):
+    """换优化器种类时不加载动量缓冲：AdamW 的 exp_avg 对 SGD 没有意义，硬塞只是白占内存。"""
+    work_dir = tmp_path / "run"
+    _make_trainer(
+        tiny_image_dir, work_dir, epochs=2,
+        optimizer={"name": "adamw", "lr": 1e-3, "weight_decay": 1e-4},
+    ).fit()
+
+    resumed = SimCLRTrainer(
+        BACKBONES.build({"name": "tiny_conv", "widths": [8, 16], "depths": [1, 1]}),
+        _make_trainer(tiny_image_dir, work_dir, epochs=1).dataloader,
+        {"epochs": 6, "amp": False, "log_interval": 999, "save_interval": 1,
+         "scheduler": {"name": "cosine"},
+         "optimizer": {"name": "sgd", "lr": 0.01, "momentum": 0.9},
+         "projection": {"hidden_channels": 16, "out_channels": 8},
+         "swanlab": {"enabled": False}},
+        torch.device("cpu"),
+        work_dir,
+        resume=work_dir / "simclr_last.pth",
+    )
+
+    assert isinstance(resumed.optimizer, torch.optim.SGD)
+    assert not resumed.optimizer.state, "不该带着 AdamW 的缓冲"
+    assert resumed.start_epoch == 2  # 训练进度仍然接上
+    assert resumed.optimizer.param_groups[0]["momentum"] == 0.9
+
+
 def test_resume_with_same_schedule_restores_state(tiny_image_dir, tmp_path):
     """总轮数不变时，重建 + 快进的结果应与原调度进度一致。"""
     work_dir = tmp_path / "run"