Prevent Arrow batch leaks when stream callbacks fail

Summary

An Arrow stream callback can report an error after placing a batch in its output parameter. The Python client currently raises the stream error immediately and skips cleanup of that batch. If the batch has a release callback, its memory and other producer-owned resources can leak.

This affects the Arrow-native DataFrame ingestion path and is most likely to appear with custom or faulty Arrow stream producers.

Why it happens

The client follows this sequence in _capsule_consume_stream:

1. Create an empty ArrowArray.
2. Pass it to ArrowArrayStream.get_next().
3. If get_next() returns a nonzero status, raise an exception.

Normal batches are released in a later try/finally block. The error branch runs before that block, so a failing callback that populated batch.release leaves the batch unreleased.

Expected behavior

• Define the ownership rule for a failed get_next() call that leaves a non-NULL batch.release callback.
• Release any producer-owned batch exactly once before reporting the stream error.
• Preserve the current behavior when the callback leaves the output array untouched.
• Keep the original stream error available to the caller.

Regression coverage

Add a small forged Arrow stream with a counting release callback and cover:

• get_next() fails without touching the output array;
• get_next() fails after populating the output array;
• the populated batch is released once, with no leak or double release.

This is separate from Arrow schema and array validation in c-questdb-client #195: the failure occurs while reading the stream, before native preflight validation begins.

Start at _capsule_consume_stream, following how ArrowArrayStream.get_next() fills an ArrowArray and how normal batches are released. Add regression coverage with a forged Arrow stream for untouched and populated output arrays. Done means a populated batch is released exactly once before the original stream error is reported, while untouched output preserves current behavior.