connectors: runtime discards the sink consume FFI status, so failed batches are counted as processed

## The shape

`iggy_sink_consume` is declared `-> i32` and the SDK returns non-zero on failure (`-1` when the batch cannot be deserialized, `1` when the sink's `consume()` errors — `sdk/src/sink.rs`). The runtime discards that status: `process_messages` calls `(consume)(...)` without binding the result and returns `Ok(SinkBatchTiming { .. })` unconditionally (`runtime/src/sink.rs`).

Since the consumer commits offsets at poll time (`AutoCommitWhen::PollingMessages`), a sink that cannot write still increments `messages_processed`, keeps `status: running` with no `last_error`, and the batch is permanently lost.

The same file checks the same channel for `iggy_sink_open` (`if result != 0 → Err`); only the per-batch call is unchecked.

## Reproduced end-to-end

postgres_source → server → postgres_sink on master, with a simulated sink DB failure: a 6-row source produces a 5-row sink. The plugin logs the insert error, the runtime still logs `Consumed 1 messages`, `/sinks` reports `running`, and the lost message is never redelivered.

## Fix

Bind the status; on non-zero, return `RuntimeError::SinkConsumeFailed { plugin_id, status }` — the caller's existing `Err` path already logs the failure and skips the processed counter. I have this ready with a regression test (stub `extern "C"` callbacks; the test fails on master with `expected SinkConsumeFailed, got success`).

---

**Affected area**: connectors runtime (`core/connectors/runtime/src/sink.rs`)
**Version**: master @ `cc269ef70`
**Logs** (from the E2E repro):
```
ERROR connector_target="iggy_connector_postgres_sink" Failed to insert batch: Cannot store data: ...
INFO iggy_connectors::sink: Consumed 1 messages in 3.9ms for sink connector with ID: 2
```

## Scope

Propagating the status is fail-fast only — it surfaces the failure but does not redeliver the batch, since offsets are already committed at poll time. The full fix is the offset-commit/DLQ design discussed in #2927.

Start in core/connectors/runtime/src/sink.rs, reading process_messages and the nearby iggy_sink_open status check. Run the connectors runtime tests, including the regression test with stub extern "C" callbacks. Done means a non-zero iggy_sink_consume status returns SinkConsumeFailed and the failed batch is not counted as processed.

Written by the indexing model from the issue text.