[Bug]: couch_peruser leaks mem3_cluster and init_changes_handler processes

### Version

CouchDB version: 3.4.3 (dpkg on Ubuntu, single-node)

### Describe the problem you're encountering

Above memory use is from an idle server - the drop off is a restart of couchdb via `systemctl restart couchdb`.

## Root cause:

init_state/0 in [couch_peruser.erl](https://github.com/apache/couchdb/blob/3.4.3/src/couch_peruser/src/couch_peruser.erl) spawns a new mem3_cluster:start_link/4 process on every cluster_unstable cast and update_config cast. exit_changes/1 cleans up change feed handlers but never terminates the previous mem3_cluster process. Old mem3_cluster processes remain alive and continue firing cluster_unstable events, each triggering another init_state/0 call, creating a feedback loop of process accumulation.

## Evidence:

Erlang remote shell inspection on the running node:

- process_count climbed from 4,097 to 4,176 in ~2 minutes on an idle system (70 HTTP requests total since last restart)
- erlang:process_info showed 16,426 couch_peruser:init_changes_handler processes accumulating
- couch_event_server (the event dispatcher) was the top memory consumer at 8.3 MB, growing as it tracked all registered handlers

## Memory timeline (from system monitoring):
- Post-restart: 3.2% RAM → 3.4% in 2 minutes
- 6:14 PM: 39.7% → 2:24 AM: 87%+ (8 hours, +5.77%/hour)
- Consistent across multiple restarts

_system snapshots (3.6 minutes apart):

```
┌───────────┬──────────────────────────┬────────────────────────────┐
│ Field │ Reading 1 (uptime 1593s) │ Reading 2 (uptime 1811s) │
├───────────┼──────────────────────────┼────────────────────────────┤
│ processes │ 63,622,424 bytes │ 70,445,232 bytes (+6.8 MB) │
├───────────┼──────────────────────────┼────────────────────────────┤
│ binary │ 787,440 │ 903,448 │
├───────────┼──────────────────────────┼────────────────────────────┤
│ ets │ 1,302,568 │ 1,308,112 │
├───────────┼──────────────────────────┼────────────────────────────┤
│ code │ 13,955,494 │ 13,955,494 │
└───────────┴──────────────────────────┴────────────────────────────┘
```

## Workaround:

Disable couch_peruser (PUT _node/_local/_config/couch_peruser/enable → "false").

### Expected Behaviour

Steady state memory consumption for an idle server.

### Steps to Reproduce

Haven't manually reproduced, but roughly:
- init 3.4.3 db
- enable couchdb per_user
- create some dbs, users
- observe

### Your Environment

- Single-node CouchDB 3.4.3 (dpkg, Ubuntu 20.04)
- 2 GB RAM
- 97 databases (including ~80 userdb-* via couch_peruser)
- Near-zero traffic
- No active tasks, smoosh idle, no replication jobs

### Additional Context

## Related prior work:

PR #3851 attempted a fix by unlinking the old mem3_cluster process, but reviewer @nickva noted it was insufficient — unlink without kill still leaves processes alive and sending events. nickva recommended keeping a single mem3_cluster process for the gen_server's lifetime and only restarting change feed handlers on state resets.

Start in src/couch_peruser/src/couch_peruser.erl by tracing init_state/0, exit_changes/1, and the mem3_cluster:start_link/4 lifecycle, then review the limitations of PR #3851. Done means couch_peruser retains the intended process lifecycle, stops accumulating init_changes_handler and mem3_cluster processes, and maintains steady memory use while idle.

Written by the indexing model from the issue text.