[Data] Fix incorrect pending task size if outputs are empty (#47604)

If an operator outputs empty blocks, then Ray Data thinks that the
operator has 256 MiB of pending task outputs, even though it should be
0. For example:
```python
import pyarrow as pa
output = pa.Table.from_pydict({"data": [None] * 128})
assert output.nbytes == 0, output.nbytes
```

The reason for the bug is because we check if `average_bytes_per_output`
is truthy rather than if it's not `None`.

https://github.com/ray-project/ray/blob/1f83fb44580e392ba6d39a9e79bbdd8cd5b7d916/python/ray/data/_internal/execution/interfaces/op_runtime_metrics.py#L369-L371
---

Signed-off-by: Balaji Veeramani <bveeramani@berkeley.edu>

python/ray/data/_internal/execution/interfaces/op_runtime_metrics.py

@@ -366,9 +366,9 @@ def obj_store_mem_max_pending_output_per_task(self) -> Optional[float]:
         if context._max_num_blocks_in_streaming_gen_buffer is None:
             return None
 
-        bytes_per_output = (
-            self.average_bytes_per_output or context.target_max_block_size
-        )
+        bytes_per_output = self.average_bytes_per_output
+        if bytes_per_output is None:
+            bytes_per_output = context.target_max_block_size
 
         num_pending_outputs = context._max_num_blocks_in_streaming_gen_buffer
         if self.average_num_outputs_per_task is not None: