Replace apex clip_grad_norm_ with PyTorch native in dints templates

apex.contrib.clip_grad.clip_grad_norm_ crashes on PyTorch >=2.10 with
"RuntimeError: Cannot access data pointer of Tensor that doesn't have
storage" because apex's multi_tensor_applier cannot handle tensors with
lazy/functional storage introduced in newer PyTorch versions.

The try/except only caught ModuleNotFoundError (apex not installed) but
not the runtime crash when apex is installed but incompatible.

torch.nn.utils.clip_grad_norm_ handles all tensor types correctly and
is the standard approach. The apex version offered marginal performance
gains that are not worth the compatibility breakage.

Fixes Project-MONAI/MONAI#8737

Author: HeyangQin

auto3dseg/algorithm_templates/dints/scripts/search.py

@@ -40,10 +40,7 @@
 from monai.metrics import compute_dice
 from monai.utils import RankFilter, set_determinism
 
-try:
-    from apex.contrib.clip_grad import clip_grad_norm_
-except ModuleNotFoundError:
-    from torch.nn.utils import clip_grad_norm_
+from torch.nn.utils import clip_grad_norm_
 
 
 CONFIG = {

auto3dseg/algorithm_templates/dints/scripts/train.py

@@ -47,10 +47,7 @@
 from torch.utils.tensorboard import SummaryWriter
 from tqdm import tqdm
 
-try:
-    from apex.contrib.clip_grad import clip_grad_norm_
-except ModuleNotFoundError:
-    from torch.nn.utils import clip_grad_norm_
+from torch.nn.utils import clip_grad_norm_
 
 try:
     _libcudart = ctypes.CDLL("libcudart.so")