bug(connectors): quickwit_sink has no timeout or retries and deadlocks on concurrent index creation

[mfyuce]

Bug description

quickwit_sink predates the shared connector retry helpers (sdk/src/retry.rs
arrived with the InfluxDB sink in #2933, roughly ten months after the sink
itself) and was never adapted to them. Compared to the other HTTP sinks it is
missing several guards, and each gap turns a recoverable condition into a hang
or into silent data loss.

1. No request timeout. create_index, has_index and ingest all share a
bare reqwest::Client::new(), which has no timeout. Under a network partition,
or against a Quickwit node that is still starting, these calls block
indefinitely. The connector neither makes progress nor reports a failure.

2. No retries. A transient 5xx or a connection reset during open() fails
the connector outright, and the same during ingest() drops the batch. Every
other HTTP sink routes these through retry middleware with exponential backoff.

3. Concurrent index creation deadlocks the connector. open() calls
has_index() and, if absent, create_index(). Two instances starting together,
or one restarting next to a live one, both observe has_index() == false and
both POST. The loser receives 409 Conflict, which is treated as a fatal
InitError, so that instance never opens even though the index it wanted now
exists.

4. Non-JSON payloads are discarded. consume() matches only
Payload::Json; everything else hits a warn! and is dropped. A topic carrying
raw or text messages silently loses every record, with the offset already
committed.

5. No readiness gate. There is no health probe before the first request, so
startup races the node coming up.

6. Trailing slash in url produces a malformed path. Every request rebuilds
its URL from self.config.url with format! and nothing trims a trailing
slash, so a configured http://host:7280/ yields //api/v1/....

Affected area / component

Connectors

Deployment

Compiled from source

Reproduction

For (3), start two quickwit_sink connectors against the same index before it
exists. One opens, the other fails with InitError on the 409 and stays down.

For (4), publish a message with a non-JSON payload to a topic the sink consumes.
The record is dropped, the offset advances, and only a warn! is emitted.

For (6), set url = "http://localhost:7280/" in [plugin_config] and observe
the request path.

Proposed scope

#3523 addresses all six:

• Build the HTTP client in open() with a configurable timeout, wrapped in
  build_retry_client from iggy_connector_sdk::retry, with configurable retry
  delay bounds.
• Probe /health/readyz through check_connectivity_with_retry before touching
  the index.
• Absorb 409 Conflict, and a 400 whose body reports the index already
  exists, as success in create_index.
• Wrap non-JSON payloads instead of dropping them: raw bytes are parsed as JSON
  when possible, otherwise carried as text, or base64 when not valid UTF-8.
• Resolve the base URL once in open() with the trailing slash trimmed.
• Add unit tests for payload handling and config defaults, alongside the
  existing integration suite in core/integration/tests/connectors/quickwit/.