[python] Add table repair fix mode and catalog integration (Part 2/3)

[JunRuiLee]

Summary

• Add TableRepair.repair(dry_run=False) to fix metadata inconsistencies
• Currently fixes: LATEST pointing to missing snapshot → rewritten to newest valid snapshot
• Add repair_table/repair_database/repair_catalog catalog API with return type annotations
• FileSystemCatalog implementation delegating to repair module
• Per-table error isolation in repair_database (continues on individual table failures)
• Fix operation is idempotent: single atomic write, safe to re-run after interruption

Context

Split from #7940 following @JingsongLi's review comment.

Depends on Part 1: #7943

• Part 1: Read-only verification logic ✅
• Part 2 (this PR): Fix mode + catalog integration
• Part 3: CLI command (depends on Part 2)

Please merge in order: Part 1 → Part 2 → Part 3.

Tests added

• test_repair_fix_dangling_latest: fix mode rewrites LATEST to newest valid snapshot
• test_dry_run_does_not_modify: dry_run=True does not change any files
• test_repair_database_level: repairs all tables in a database
• test_repair_catalog_level: repairs all tables across all databases
• test_fix_latest_respects_check_data_files: fix skips snapshots with missing data files
• test_repair_is_idempotent: running repair twice converges — second run is a no-op

[JingsongLi]

Good follow-up. The fix mode and catalog integration look reasonable. Some feedback:

1. _fix_latest_file atomicity: The fallback path (delete + overwrite when try_to_write_atomic returns False) has a small window where LATEST doesn't exist. In the Java Paimon implementation, SnapshotManager uses an overwrite-in-place strategy for hint files since they're best-effort. Consider just doing overwrite_file_utf8 directly without the delete step — the atomic write is a nice-to-have for LATEST but not strictly required.

2. repair_database / repair_catalog filtering: The directory listing uses name.endswith(".db") to find databases. This could accidentally pick up non-database directories that happen to end with .db. The Java catalog uses a metadata-based approach (listing from catalog metadata). Since this is filesystem-only, it's acceptable, but worth a comment noting the limitation.

3. except DatabaseNotExistException: raise in filesystem_catalog.py — this try/except that just re-raises is a no-op. You can simplify to just call self.get_database(database_name) without the try block.

4. Test duplication across Part 1 and Part 2: The test file in this PR (940 lines) includes all tests from Part 1 plus the new fix-mode tests. Since Part 2 depends on Part 1, the test file should only contain the new TestTableRepairFixMode class, with Part 1's tests already merged. Otherwise there will be conflicts when merging.

5. Missing check_data_files in CLI: The CLI cmd_table_repair doesn't expose --check-data-files flag. Should it? Users who want to do a deep check would need to use the Python API directly.

[JunRuiLee]

Thanks for the review! Marking this as Draft since it depends on Part 1 (#7943). Once #7943 is merged, I'll rebase onto master so the diff only contains the Part 2 incremental changes (fix mode + catalog integration). Will address your feedback points at that time.