fix: persist mount metadata across master switch

[jlon]

Summary

• flush mount/unmount journal entries before returning success to avoid mount metadata loss during leader switch
• fix mount update-mode existence check to use curvine mount-path index
• expose standby master fs in mini cluster test helper for failover assertions
• add mount regression and coverage tests in curvine-tests (failover, mount manager, mount table)

Verification

• cargo clippy --all-targets --jobs 2 -- --deny=warnings --allow clippy::uninlined-format-args
• cargo test -p curvine-tests --test mount_failover_test --test mount_manager_test --test mount_table_test -- --nocapture

[jlon]

Additional note for this PR update (commit: d340099):

Why this change

After fixing the mount-loss issue, we also need clear diagnostics for MountTable::restore() so that restore-time failures are visible instead of silent.

What was added

File: curvine-server/src/master/mount/mount_table.rs

• Failed to load mount table from metadata store:
  • mount restore failed: unable to load mount table from metadata store, err=...
• Restore start with total entries:
  • mount restore started: <N> entries loaded from metadata store
• Empty-table case:
  • mount restore completed: no entries found
• Per-entry success:
  • mount restore entry succeeded: mount_id=..., cv_path=..., ufs_path=...
• Per-entry failure:
  • mount restore entry failed: mount_id=..., cv_path=..., ufs_path=..., err=...
• Final summary:
  • all success: mount restore completed successfully: restored=..., failed=0
  • partial failure: mount restore completed with errors: restored=..., failed=...

Behavioral scope

• This is observability-only for restore path.
• No fail-fast behavior change was introduced in this commit.

Validation

• mount_table_test
• mount_manager_test
• mount_failover_test
  All passed.

[szbr9486]

else {
info!(
"mount restore started: {} entries loaded from metadata store",
total
);
}

[jlon]

Keeping this outside else is intentional. We want a clear restore lifecycle log even when metadata load succeeds with zero entries, so operators can distinguish "restore executed and found 0 mounts" from "restore never ran".

[jlon]

Updated in 0b5a30e. The start log now runs only when total > 0, so the empty-table path emits only a single completion line.

[szbr9486]

delete

[jlon]

total is kept to support restore observability and summary correlation. It is used to report how many entries were loaded from metadata before per-entry restore, which helps diagnose partial-restore cases.

[jlon]

Updated in 0b5a30e. Removed redundant temp vars in the restore loop and simplified logging flow.

[szbr9486]

Deleting the three lines above only prints in the logs, so it's unnecessary.

[jlon]

I am keeping these logs on purpose. They provide low-cost but important startup visibility for master failover: load success/failure, empty-table case, and restore progress are distinct states in production troubleshooting.

[jlon]

Updated in 0b5a30e. Empty restore now logs once (no entries found) and avoids duplicate-looking startup lines.