fix(examples): Avoid torch dataset formatter

[WilliamLindskog]

What changed

• Remove Dataset.with_format("torch") from the PyTorch example paths that combine Hugging Face Datasets with torchvision transforms.
• Add a shared lazy tensor transform for the Whisper example so encoded data and targets are still returned as PyTorch tensors without using the global torch formatter.

Why

This is a follow-up to #7330. The E2E blocker was fixed there, but a few examples still used with_format("torch"), which can trigger Hugging Face Datasets' torch formatter import path and the unsupported torchvision.io.VideoReader import under some dependency combinations.

This PR is intentionally examples-only and does not change package minimum versions.

Validation

• ruff, black --check, and py_compile on touched example Python files
• git diff --check

[Copilot]

Pull request overview

This PR removes usage of Hugging Face Datasets’ global Torch formatter (Dataset.with_format("torch")) from several PyTorch-facing examples and docs to avoid triggering the fragile torchvision.io.VideoReader import path under certain dependency combinations, while preserving tensor outputs via lazy transforms where needed.

Changes:

• Replaced with_format("torch") with with_transform(...) in CIFAR10-based PyTorch examples (quickstart + custom-mods).
• Introduced a shared lazy tensor conversion transform for the Whisper example and applied it across centralized/client/server evaluation paths.
• Removed with_format("torch") from the FedDebug label-noise dataset path and updated the PyTorch guide to stop recommending the map(...).with_format("torch") pattern.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
examples/whisper-federated-finetuning/whisper_example/server_app.py	Uses the new lazy torch transform for central evaluation datasets instead of with_format("torch").
examples/whisper-federated-finetuning/whisper_example/dataset.py	Adds with_torch_transform helper (lazy tensor conversion for encoded data/targets).
examples/whisper-federated-finetuning/whisper_example/client_app.py	Applies lazy torch transform to training partition after sampler construction; removes with_format("torch").
examples/whisper-federated-finetuning/centralized.py	Switches centralized train/val/test datasets to the lazy torch transform helper.
examples/quickstart-pytorch/pytorchexample/task.py	Removes global torch formatter from centralized CIFAR10 DataLoader path.
examples/custom-mods/custom_mods/task.py	Removes global torch formatter from centralized CIFAR10 DataLoader path.
datasets/docs/source/how-to-use-with-pytorch.rst	Removes documentation recommending map(...).with_format("torch").
baselines/feddebug/feddebug/dataset.py	Drops with_format("torch") from label-noise mapping path.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[WilliamLindskog]

Addressing the scope/dependency concern explicitly:

• I narrowed this PR to examples-only. It now only touches the PyTorch/custom-mods/Whisper example files; the docs and FedDebug baseline changes were removed.
• I do not think this needs a minimum torch/torchvision version bump. The PR removes Dataset.with_format("torch"); it does not rely on a newer torch/torchvision API. The replacement paths use existing example transforms and, for Whisper, torch.as_tensor to preserve tensor output for data/targets.
• The motivation is to avoid the Hugging Face Datasets torch formatter import path that can hit torchvision.io.VideoReader under some dependency combinations, while keeping the example behavior equivalent.