sqlinstance: use the correct timestamp when populating the cache

Previously, the sql instance cache would be populated using the
timestamp of the initial rangefeed scan. This was a bug since the
timestamp is used to determine which sql instance ID to use when there
are two IDs for the same address (i.e., it should use the most recent
one).

This caused problems if a tenant node was shutdown, then started with
the same address. Other layers like DistSQL would try to route requests
to an old sql instance, which would actually be the same as the
requesting node.

No release note since this bug didn't affect anything prior to v22.2.

Release justification: high priority bug fix.

pkg/sql/sqlinstance/instancestorage/instancereader.go

@@ -162,6 +162,7 @@ func (r *Reader) maybeStartRangeFeed(ctx context.Context) *rangefeed.RangeFeed {
 		updateCacheFn,
 		rangefeed.WithInitialScan(initialScanDoneFn),
 		rangefeed.WithOnInitialScanError(initialScanErrFn),
+		rangefeed.WithRowTimestampInInitialScan(true),
 	)
 	r.setStarted()
 	if err != nil {