Expand 3D captcha into three subtypes: 3d_text, 3d_rotate, 3d_slider

Split the single "3d" captcha type into three independent expert models:
- 3d_text: 3D perspective text OCR (renamed from old "3d", CTC-based ThreeDCNN)
- 3d_rotate: rotation angle regression (new RegressionCNN, circular loss)
- 3d_slider: slider offset regression (new RegressionCNN, SmoothL1 loss)

CAPTCHA_TYPES expanded from 3 to 5 classes. Classifier samples updated
to 50000 (10000 per class). New generators, model, dataset, training
utilities, and full pipeline/export/CLI support for all subtypes.

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

models/regression_cnn.py

@@ -0,0 +1,86 @@
+"""
+回归 CNN 模型
+
+3d_rotate 和 3d_slider 共用的回归模型。
+输出 sigmoid 归一化到 [0,1]，推理时按 label_range 缩放回原始范围。
+
+架构:
+    Conv(1→32) + BN + ReLU + Pool
+    Conv(32→64) + BN + ReLU + Pool
+    Conv(64→128) + BN + ReLU + Pool
+    Conv(128→128) + BN + ReLU + Pool
+    AdaptiveAvgPool2d(1) → FC(128→64) → ReLU → Dropout(0.2) → FC(64→1) → Sigmoid
+
+约 250K 参数，~1MB。
+"""
+
+import torch
+import torch.nn as nn
+
+
+class RegressionCNN(nn.Module):
+    """
+    轻量回归 CNN，用于 3d_rotate (角度) 和 3d_slider (偏移) 预测。
+
+    输出 [0, 1] 范围的 sigmoid 值，需要按 label_range 缩放到实际范围。
+    """
+
+    def __init__(self, img_h: int = 80, img_w: int = 80):
+        """
+        Args:
+            img_h: 输入图片高度
+            img_w: 输入图片宽度
+        """
+        super().__init__()
+        self.img_h = img_h
+        self.img_w = img_w
+
+        self.features = nn.Sequential(
+            # block 1: 1 → 32, H/2, W/2
+            nn.Conv2d(1, 32, kernel_size=3, padding=1, bias=False),
+            nn.BatchNorm2d(32),
+            nn.ReLU(inplace=True),
+            nn.MaxPool2d(2, 2),
+
+            # block 2: 32 → 64, H/4, W/4
+            nn.Conv2d(32, 64, kernel_size=3, padding=1, bias=False),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True),
+            nn.MaxPool2d(2, 2),
+
+            # block 3: 64 → 128, H/8, W/8
+            nn.Conv2d(64, 128, kernel_size=3, padding=1, bias=False),
+            nn.BatchNorm2d(128),
+            nn.ReLU(inplace=True),
+            nn.MaxPool2d(2, 2),
+
+            # block 4: 128 → 128, H/16, W/16
+            nn.Conv2d(128, 128, kernel_size=3, padding=1, bias=False),
+            nn.BatchNorm2d(128),
+            nn.ReLU(inplace=True),
+            nn.MaxPool2d(2, 2),
+        )
+
+        self.pool = nn.AdaptiveAvgPool2d(1)
+
+        self.regressor = nn.Sequential(
+            nn.Linear(128, 64),
+            nn.ReLU(inplace=True),
+            nn.Dropout(0.2),
+            nn.Linear(64, 1),
+            nn.Sigmoid(),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            x: (batch, 1, H, W) 灰度图
+
+        Returns:
+            output: (batch, 1) sigmoid 输出 [0, 1]
+        """
+        feat = self.features(x)
+        feat = self.pool(feat)          # (B, 128, 1, 1)
+        feat = feat.flatten(1)          # (B, 128)
+        out = self.regressor(feat)      # (B, 1)
+        return out