…_type`

- Rework channel source to maintain a pubsub session that auto-reconnects
  and re-subscribes after Redis disconnects.
- Add shutdown-aware backoff and best-effort unsubscribe on exit.
- Log when the connection is re-established ("re-established and resubscribed").

Resolves vectordotdev#22615

- Race `connect_and_subscribe` with the shutdown signal (biased select) so a
  connect against a black-holed endpoint can't block graceful shutdown until
  the force-shutdown deadline.
- Drop the explicit `UNSUBSCRIBE` round trips on the shutdown/disconnect paths;
  dropping the pubsub connection closes it and Redis releases the subscription
  automatically, so shutdown no longer waits on a network round trip.
- Keep the `list` source's retry backoff capped at 1s (its prior behavior)
  rather than the channel source's 30s cap, avoiding a recovery-latency
  regression for existing `data_type = "list"` users.

Second round of automated review feedback:

- Do the initial connect + SUBSCRIBE during source build and fail fast on
  non-recoverable errors (bad auth/ACL/config) by returning the error, instead
  of entering the reconnect loop. An invalid configuration now surfaces at
  startup rather than looking like it started and only erroring at runtime.
  Transient (I/O) errors are still tolerated and handled by the reconnect loop.
- Only reset the reconnect backoff after a session has actually delivered a
  message, not merely on a successful connect. A connection that is accepted and
  then immediately dropped (a flapping outage) now keeps backing off
  exponentially instead of hot-looping at the 500ms floor.
- Classify permanent vs transient connect errors during reconnect too: a
  non-recoverable error stops the source rather than retrying forever.

Use redis-rs's own `RetryMethod` classification instead of only treating
`ErrorKind::IoError` as transient. A Redis node coming back from a restart may
accept the TCP connection but reply to `SUBSCRIBE` with a server-side retryable
error such as `LOADING`/`BUSYLOADING`/`TRYAGAIN`, which redis-rs reports as a
non-I/O error. The previous check treated those as permanent and stopped the
source, defeating the restart recovery this change is meant to provide. Now only
`RetryMethod::NoRetry` (auth/ACL/config rejections) is fatal; all other errors,
including those retryable server states, back off and reconnect.

… sessions

Third round of automated review feedback:

- redis-rs maps `ErrorKind::AuthenticationFailed` to `RetryMethod::Reconnect`,
  so the `RetryMethod`-based classifier treated bad credentials as transient and
  retried forever. Treat `AuthenticationFailed` as permanent explicitly so an
  invalid password/ACL fails the source build as intended.
- Reset the reconnect backoff after a session has either delivered a message or
  stayed connected for a minimum healthy duration, rather than only on delivered
  data. A low-volume channel that stays healthy but quiet now clears a backoff
  that a previous flapping period drove up to the cap, so a later drop resubscribes
  promptly instead of waiting up to 30s.

Replace the error-classification approach with the source's original build-time
semantics, which is both simpler and more robust:

- The initial connect + SUBSCRIBE fails fast on ANY error (auth, TLS, ACL,
  unreachable, LOADING), surfacing misconfiguration at startup as the source did
  before reconnect support. This avoids fragile transient-vs-permanent
  classification: redis-rs reports TLS handshake failures and transient network
  timeouts both as IoError/RetryImmediately, so no kind/retry_method predicate can
  separate them. Removing `is_transient` also drops the auth special-case.
- Once the source is running, every reconnect failure is retried; a permanent
  problem was already ruled out by the successful build-time connect. Failures are
  still recorded via RedisConnectionError for alerting.
- Emit a new RedisConnectionDropped internal event (component_errors_total) when an
  established connection drops, so alerts fire even when the reconnect succeeds on
  the first attempt.

`vdev check events` requires any internal event that increments
`component_errors_total` to be named `*Error`. Rename
`RedisConnectionDropped` -> `RedisConnectionDroppedError`.