volume watcher should back off if plugin unavailable

While testing out some other issues, I discovered that if a plugin is completely removed but there are allocations with claims, the volumewatcher gets into a tight error loop. This only seems to happen if there are terminal allocations with claims have not yet been reaped. Everything recovers as expected if the plugin is re-registered.

> 2023-03-30T11:39:09.016-0400 [ERROR] nomad.volumes_watcher: error releasing volume claims: namespace=default volume_id=volume1
> error=
> | 1 error occurred:
> | \t* no such plugin: "hostpath-plugin0"
> |

(Also these are ugly multi-errors when they don't need to be. Would be nice to fix that while we're in there.)

Start by locating the volume watcher and reproduce the tight error loop with a removed plugin, terminal allocations, and unreaped claims. Verify the behavior when the plugin is re-registered, then make completion mean that unavailable plugins are retried with backoff and the unnecessary multi-error output is cleaned up.

Written by the indexing model from the issue text.