schemastore: lazy keyspace initialization can block graceful shutdown

What did you do?

Run the NextGen storage-sink fail_over integration test with two TiCDC
nodes. The test creates a changefeed for a non-default keyspace, terminates one
node, starts its replacement, and then terminates the other node while the new
coordinator is bootstrapping the changefeed.

The failure was observed in
pull-cdc-storage-integration-heavy-next-gen/723,
storage group G13.

What did you expect to see?

SIGTERM should cancel any in-flight lazy schema-store initialization. The TiCDC
process should close within the graceful-shutdown deadline, allowing failover
or scale-in automation to start a replacement safely.

What did you see instead?

The process accepted SIGTERM but did not exit within 30 seconds:

[18:49:29.313] got signal, prepare to shutdown
[18:49:29.316] schema store initialize from kv storage begin
[18:49:29-18:49:58] loadRegion from PD failed ... ErrClientGetProtoClient
[18:49:59.315] graceful shutdown timeout, exit server
wait process 4983 exit timeout

The shutdown race is:

1. getKeyspaceSchemaStore starts lazy registration using
   context.Background().
2. RegisterKeyspace holds keyspaceLocker while constructing the persistent
   schema store.
3. persistSchemaSnapshotWithEncryption retries ListDatabases indefinitely
   and does not observe a cancellation context.
4. During shutdown, the PD client is closed and the snapshot initialization
   keeps retrying. SchemaStore.Close waits for keyspaceLocker, so the server
   reaches its graceful-shutdown timeout.

This can make graceful scale-in or failover report an error and leave the old
process alive until an external hard-kill deadline. No data inconsistency was
observed in this run; the test stopped before Sync Diff.

Suggested direction:

• propagate a server-lifecycle context into lazy keyspace registration;
• make schema snapshot retry loops cancellation-aware; and
• avoid holding the global keyspace lock across unbounded I/O, or otherwise
  allow Close to cancel and wait for in-flight initialization.

Versions of the cluster

Upstream TiDB cluster version:

CLOUD.202609.0-10f06594e7

Upstream TiKV version:

Not captured

TiCDC version:

v8.5.4-nextgen.202510.5-417-g66509e02
Presubmit merge: base 6c03227f9 + head 339b9b66a
Kernel: NextGen

Start at getKeyspaceSchemaStore and RegisterKeyspace, then trace persistSchemaSnapshotWithEncryption and SchemaStore.Close. Reproduce the two-node fail_over integration scenario and follow shutdown and retry behavior around the keyspace lock. Done means SIGTERM cancels lazy schema-store initialization and the process exits within the graceful-shutdown deadline.