Zombie agent trapped in catalog

When filing a bug, please include the following headings if possible. Any example text in this template can be deleted.

#### Overview of the Issue

I am seeing a lot of errors like this in my server logs:

2021-08-13T10:43:35.033Z [INFO] agent.server.snapshot: creating new snapshot: path=data\raft\snapshots\101973-54331185-1628851415033.tmp
2021-08-13T10:43:35.050Z [INFO] agent.server.snapshot: reaping snapshot: path=data\raft\snapshots\101973-54331185-1628851394441
2021-08-13T10:43:35.051Z [INFO] agent.server.raft: copied to local snapshot: bytes=579416
2021-08-13T10:43:35.055Z [ERROR] agent.server.raft: failed to restore snapshot: error="failed to restore snapshot 101973-54331185-1628851415033: check node "EW1A-TST-MPC" does not match node "ew1a-tst-mpc""
2021-08-13T10:43:35.862Z [ERROR] agent.anti_entropy: failed to sync remote state: error="ACL not found"

and on others:

Aug 13 10:08:05 ip-10-142-20-109 consul[1944]: 2021-08-13T10:08:05.944Z [WARN] agent.fsm: DeleteNode failed: error="failed generating events from changes: Missing node registration"

On agents I see:

2021-08-13T10:43:35.862Z [ERROR] agent.anti_entropy: failed to sync remote state: error="ACL not found"

The errors and warnings just repeat. I have tried force-leave and -prune but they report the node doesnt exist.

#### Reproduction Steps

I believe what has happened is:

- a node registered itself EW1A-TST-MPC
- this should be lower case so it was re-registered via config with node-name ew1a-tst-mpc
- that instance was then destroyed
- in the period before consul would have cleaned it up, its agent token was removed
- when consul went to remove the node it wasnt able to complete the process

Now its left in this "zombie" state where its too dead for force-leave to work but not dead enough to get cleaned up

### Consul info for both Client and Server

1.9,8 windows and linux servers

No source file or test is named. Start by examining the Consul agent logs around snapshot restoration, anti-entropy ACL sync, and FSM DeleteNode failures, then reproduce the sequence involving case-changed node names and force-leave or -prune. Done means the zombie node is cleaned up and the repeated errors stop.

Written by the indexing model from the issue text.