Fix infinite retry loops, bare asserts, double iteration counter, and stale docstring

[Copilot]

Several reliability and correctness bugs across the data loading and training pipeline. Key issues: retry loops with no exit condition could hang indefinitely, bare assert statements are silently stripped under -O, and self.iteration was double-incremented per training step.

Fixes

• Infinite retry loops (data_loader.py, features_loader.py): All three while not succ retry loops now accept max_retries=10 and raise RuntimeError after exhaustion instead of spinning forever
• Bare asserts → FileNotFoundError (features_loader.py): assert os.path.exists(path) replaced with explicit guards in both FeaturesLoader._get_features_list and FeaturesLoaderVal._get_features_list
• Double self.iteration increment (network/TorchUtils.py): self.iteration += 1 appeared at both the top and bottom of the do_epoch loop body, advancing the counter by 2 per step — removed the duplicate at the bottom
• Placeholder docstring (features_loader.py): get_feature returned _type_: _description_ — replaced with accurate return type annotation tuple[Tensor, Tensor, int] and description
• Misleading log prefixes: FeaturesLoader.__getitem__ and FeaturesLoaderVal.__getitem__ logged errors as VideoIter:: — corrected to their actual class names

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.