Setting `"train"` to `classes` argument of `LSUN()` kills the kernel on JupyterLab

🐛 Describe the bug

Setting "train" to classes argument of LSUN() kills the kernel on JupyterLab as shown below:
*Memos:

• It also happens with ["bedroom_train"], ["church_outdoor_train"], ["classroom_train"], ["kitchen_train"], ["living_room_train"] and ["restaurant_train"] for classes argument.
• "test" doesn't work because test_lmdb.zip, test_train_lmdb.zip and test_val_lmdb.zip cannot be unzipped as I reported the bug.
• There is other error as I reported here.

from torchvision import datasets

train_data = datasets.LSUN(
    root="C:/Users/<user>/dl-models/data/lsun",
    classes="train" # Here
) # Error

print(len(train_data))

In addition, setting "val" to classes argument of LSUN() works properly on JupyterLab as shown below.
*Memos:

• ["bedroom_train_val"], ["bridge_val"], ["church_outdoor_val"], ["classroom_val"], ["conference_room_val"], ["dining_room_val"], ["kitchen_val"], ["living_room_val"], ["restaurant_val"] and ["tower_val"] for classes argument also work properly.
• test doesn't work because test_lmdb.zip, test_train_lmdb.zip and test_val_lmdb.zip cannot be unzipped as I reported the bug.

from torchvision import datasets

val_data = datasets.LSUN(
    root="C:/Users/<user>/dl-models/data/lsun",
    classes="val" # Here
)

print(len(val_data)) # 300

Versions

import torchvision

torchvision.__version__ # '0.20.1'

Reproduce the failure with torchvision 0.20.1 by constructing LSUN with classes="train" in JupyterLab, then compare it with the working classes="val" case. The issue names no source file or test, so trace LSUN dataset initialization and add a regression check once the training dataset no longer kills the kernel.