Classification immediate fixes and enhancements

[nishika26]

This issue covers only the points listed below. All other To-Dos from the parent issue are out of scope for this release but definitely will be taken up very soon after we ship this version of classification feature to production.

Tasks in scope (scope as of Sept 1 - Sept 2)

1. Error handling in background jobs

   • For invalid API key errors (and any other background job errors), the exact error message should be persisted in the database.

   • Errors must not only appear in logs — they should also be available in the DB, word-for-word, via the error column.

2. Update api/v1/fine_tuning/{job_id}/refresh endpoint

   • Replace job_id with id to avoid confusion with provider_job_id.

   • Exclude deleted_at from the response.

3. Initialization changes

   • Shift storage and DocumentCrud initialization to use project_id instead of user_id, since PR #346 has been merged.

4. Signed URLs for S3 files

   • In API responses, return signed S3 URLs (clickable) instead of raw s3:// URIs.

5. Testing improvements

   • Mock Session in test cases, because router and background jobs run with separate sessions.

   • This mocking is required to ensure tests work correctly with the transactional test DB.