Use cluster instead of worker_threads

## Cluster

Pro:
- Load balancing happens by kernel/Node.js distributing UDP messages across all sockets listening on the same host:port.
- No per-query IPC.
- Scales horizontally even with a single DoH upstream.
- Simpler design. Every cluster worker is an independent "master." Without the message passing and state management between threaded master/workers.

Con:
- One HTTP/2 session per origin per worker. Potentially less efficient on the network than re-using the same session.

## Worker Threads

Pro:
- Single HTTP/2 session per DoH resolver.
- Shiny new API.

Con:
- Wasted IPC on single-core CPU of Raspberry Pi or restricted Docker.
- Fixed overhead of IPC between worker/master.
- Cloning buffers is slower than expected. Sharing memory may be better but hard to measure performance (no support yet in Clinic.js).