Handle PIL UnidentifiedImageError exception when running cleanvision on local image folder dataset

[pavansai018]

Summary

This PR ensures that corrupted / unreadable images are filtered out at dataset construction time, so they never appear in dataset.index.

This prevents runtime crashes in downstream code paths such as find_issues() and visualize(), which assume that every index corresponds to a readable image.

Fixes #222

Motivation

Currently, dataset indices are created purely from discovered filepaths (or integer indices for torchvision datasets), without checking whether the underlying image data is actually readable.

As a result:

• Corrupted images can enter dataset.index
• Errors surface later during visualization or issue detection
• Failures occur far away from the root cause and are hard to recover from

Since visualize() and issue managers materialize indices before accessing images, lazy handling in __getitem__ is insufficient.

The correct place to handle this is before the index is finalized.

---

What this PR changes

File-based datasets (FSDataset)

• During filepath discovery, each image is opened once to check integrity
• Corrupted images are silently skipped
• Only valid image paths are included in _filepaths and dataset.index

TorchVision datasets (TorchDataset)

• Dataset indices are validated once during _set_index()
• Any sample whose image cannot be accessed or is not a valid PIL.Image is excluded
• dataset.index contains only readable samples

---

What this PR does NOT do

• ❌ No changes to visualize() or issue managers
• ❌ No None propagation or sentinel values
• ❌ No behavior changes for valid datasets

All downstream code continues to rely on the existing invariant:

Every index in dataset.index maps to a valid image

---

Performance considerations

• Each image/sample is checked once at dataset construction time
• This is unavoidable if dataset membership must exclude corrupted entries
• For valid images, there is no additional overhead during processing or visualization

---

Result

• dataset.index is always consistent
• Visualization never crashes due to corrupted images
• Errors are handled at the correct architectural boundary

This makes dataset handling more robust while keeping the rest of the codebase unchanged.

[CLAassistant]

All committers have signed the CLA.