Top-level listener supervisor.

rest_for_one over three children:

1. nquic_listener_mgr: gen_server with the accept queue, parked acceptors, resolved port, static key, and a published reference to the listener's dispatch table.
2. nquic_partitions_sup: one_for_one container holding the per- scheduler nquic_server_sup partitions. Each partition publishes its pid into the dispatch table on init.
3. nquic_receiver_sup: one_for_one container of N nquic_receiver workers (sharing the listen port via SO_REUSEPORT when N > 1).

The dispatch ETS table is created here, owned by this supervisor process, and lives for the lifetime of the listener. Children publish their own pids into it on init; nothing else has to coordinate restart fan-out.

On rest_for_one the cascade is:

• mgr crash -> mgr + partitions + receivers all restart; fresh dispatch slots are populated by each new child's init.
• partitions crash -> partitions + receivers restart.
• receiver_sup crash -> only receivers restart.

A single partition crashing is contained inside nquic_partitions_sup (one_for_one there) and re-publishes its slot on its own restart.