rpc: ignore offline/unassigned regions during cache warmup

When warming up the cache, ignore regions that are offline/unassigned.
They could exist for many reasons:
- parent region after a split, until daughter regions are compacted
- regions that are merged together, until compaction happens
- region moving due to HBase LoadBalancer, RegionServer shutting down or
  manually
- etc

It's not critical if we are missing a few regions in the cache during
warmup. The most important thing is for the warmup to be fast and to
fill in with as many regions as possible. Other regions will be fetched
later as needed.

rpc.go

@@ -757,7 +757,7 @@ func createAllRegionSearchKey(table []byte) []byte {
 	return metaKey
 }
 
-// metaLookupForTable checks meta table for all the region in which the given table is.
+// metaLookupForTable checks meta table for all the open table regions.
 func (c *client) metaLookupForTable(ctx context.Context,
 	table []byte) ([]regionInfoAndAddr, error) {
 	metaKey := createAllRegionSearchKey(table)
@@ -780,7 +780,16 @@ func (c *client) metaLookupForTable(ctx context.Context,
 
 		reg, addr, err := region.ParseRegionInfo(resp)
 		if err != nil {
-			return nil, err
+			// Ignore error, but log if it's anything else than OfflineRegionError. This really
+			// shouldn't happen unless HBase meta table is corrupted/changed format.
+			if _, ok := err.(region.OfflineRegionError); !ok {
+				c.logger.Debug("failed to parse region", "err", err)
+			}
+			continue
+		}
+		if addr == "" {
+			// Ignore region that aren't open/assigned to a RegionServer.
+			continue
 		}
 
 		regions = append(regions, regionInfoAndAddr{regionInfo: reg, addr: addr})