Support checkpoint for ingest mode of adding index

[tangenta]

Enhancement

Currently, TiDB doesn't support the checkpoint when using ingest mode for adding index. It means if TiDB is killed for upgrade or restart, the ongoing adding index progress is lost. This is obviously unacceptable for tasks that take a long time to complete.

The ingest mode is a better backfilling implementation introduced by TiDB v.6.3.0. TiDB uses the interface provided by tidb-lightning to write index data to the local storage, and imports the locally stored SST file. Since this implementation bypasses the transaction layer, it lacks the persistence provided by transactions. We need to implement checkpoint functions separately.

Here is the goal: for the same instance, the restarted TiDB can retain the adding index work progress and continue to execute.

Note that we don't support the resumption for the following cases:

• In a distributed environment, TiDB hangs up and the task executor (DDL owner) changes.
• The local temporary file is corrupted.