fix: improve LFW dataset download reliability with fallback URLs (#267)

[ANAMASGARD]

FIXES #267

• LFW dataset downloads now have improved reliability through fallback URL mechanism and retry logic.

Changes

• Added download_with_fallback() helper function in R/utils.R that:

  • Tries multiple mirror URLs in sequence
  • Implements configurable retry logic (default: 2 retries per URL)
  • Verifies MD5 checksums when provided
  • Provides informative error messages with manual download instructions when all mirrors fail

• Updated lfw_people_dataset and lfw_pairs_dataset in R/dataset-lfw.R:

  • Restructured URL configuration to support multiple mirrors (base_urls list)
  • Updated resource definitions with structured format (file_ids, md5)
  • Modified download() method to use the new fallback mechanism

• Added tests/testthat/test-lfw-reliability.R with tests for:

  • Function existence and behavior
  • Error message quality
  • Dataset constructor validation

Testing

• All existing LFW tests continue to pass
• New reliability tests pass
• Package builds and checks successfully

Notes

• Currently uses Figshare as the primary mirror
• Architecture supports adding additional mirrors in future if identified
• Maintains backward compatibility with existing dataset API

[cregouby]

Hello @ANAMASGARD,

Thanks for the proposal.

Your provided alternative URL is not opened, as I get a trying to open it :

<Error>
<Code>AccessDenied</Code>
<Message>Access denied.</Message>
<Details>Anonymous caller does not have storage.objects.get access to the Google Cloud Storage object. Permission 'storage.objects.get' denied on resource (or it may not exist).</Details>
</Error>

Which means switching to download_with_fallback() might be of no use...

[ANAMASGARD]

Thanks for the feedback @cregouby! You're right - I've removed the non-functional Kaggle URL in the latest commit.

The main value of this PR is now focused on improving download reliability for the existing Figshare URL:

• Retry logic - Each URL is tried 2 times with a configurable delay before failing (handles transient network issues like timeouts that were causing CI failures)
• Better error messages - When downloads fail, users get:

1. List of URLs that were attempted
2. Clear instructions for manual download with the cache directory path
3. The actual error message for debugging

• MD5 verification - Validates downloaded files aren't corrupted before extraction
• Future-proof infrastructure - The base_urls list makes it easy to add additional mirrors when reliable ones are identified
  Even with a single mirror, the retry mechanism should help reduce the intermittent download failures that were causing test flakiness. Would you like me to make any adjustments to this approach?