Normalize the gRPC error handling

As #5161 and tikv/tikv#12345 show, the gRPC error handling of PD is arbitrary. We mix using the gRPC error code and response header to return the error, which causes the TiKV may reconnect the PD incorrectly and unnecessarily. This is the tracking issue for normalizing the gRPC error handling of PD.

## TODOs

- [x] #5161
- [x] Normalize the definition of the two types of error on the PD side.
- gRPC Connection Error: an error that happens on the gRPC connection level, e.g, not the leader or unmatched cluster ID, which should cause the TiKV to re-establish the connection.
- gRPC Interface Error: an error that happens inside the gRPC interface logic, e.g, region/store not found or other PD internal errors, which should cause the TiKV to retry rather than reconnect.
- [ ] Redefine the error proto https://github.com/pingcap/kvproto/blob/08b02befd813c6f5e6629fc24ab249da98e87bb3/proto/pdpb.proto#L142
- [ ] Normalize the gRPC error handling on PD.
- [ ] Normalize the gRPC error handling on TiKV.