refactor(benchmark): delegate model loading to tasks and support CIFAR-100

- Extract model loading logic from benchmark CLI into task-owned prepare_benchmark
- Add RetrievalEncoder class wrapping DINO with optional hash compression
- Add accelerate dependency for device management  
- Switch dataset from CIFAR-10 to CIFAR-100 with fine_label column

mini-nav/utils/feature_extractor.py

@@ -3,11 +3,13 @@
 from typing import Any, List, Union, cast
 
 import torch
+from aiohttp.web import get
 from PIL import Image
+from rich.progress import track
 from torch import nn
 from torch.utils.data import DataLoader
 from transformers import BitImageProcessor
-from rich.progress import track
+from utils import get_device
 
 
 def _extract_features_from_output(output: Any) -> torch.Tensor:
@@ -26,6 +28,7 @@ def _extract_features_from_output(output: Any) -> torch.Tensor:
     return cast(torch.Tensor, output)
 
 
+@torch.no_grad()
 def infer_vector_dim(
     processor: BitImageProcessor,
     model: nn.Module,
@@ -41,13 +44,9 @@ def infer_vector_dim(
     Returns:
         Vector dimension.
     """
-    device = next(model.parameters()).device
-    model.eval()
-
-    with torch.no_grad():
-        inputs = processor(images=sample_image, return_tensors="pt")
-        inputs.to(device)
-        output = model(inputs)
+    inputs = processor(images=sample_image, return_tensors="pt")
+    inputs.to(get_device())
+    output = model(inputs)
 
     features = _extract_features_from_output(output)
     return features.shape[-1]
@@ -69,11 +68,8 @@ def extract_single_image_feature(
     Returns:
         The extracted CLS token feature vector as a list of floats.
     """
-    device = next(model.parameters()).device
-    model.eval()
-
     inputs = processor(images=image, return_tensors="pt")
-    inputs.to(device, non_blocking=True)
+    inputs.to(get_device(), non_blocking=True)
     outputs = model(inputs)
 
     features = _extract_features_from_output(outputs)  # [1, D]
@@ -98,8 +94,7 @@ def extract_batch_features(
     Returns:
         Tensor of shape [batch_size, feature_dim].
     """
-    device = next(model.parameters()).device
-    model.eval()
+    device = get_device()
 
     # Handle DataLoader input
     if isinstance(images, DataLoader):