Add IO runtime support to IcebergPartitionedScan

[toutane]

Summary

This PR implements Option 3 as described in this document and is linked to Jira ticket QECO-1015.

The upstream PR apache#2308 is also related to this feature and would replace the proposed implementation when merged if relevant.

Adds optional IO runtime segregation to IcebergPartitionedScan and IcebergPartitionedTableProvider, mirroring the IOExec pattern from dd-datafusion.

What

IcebergPartitionedScan

• New optional io_handle: Option<tokio::runtime::Handle> field, set via with_io_handle(Handle).
• When set, execute() spawns the Parquet read (opendal) on the IO runtime and bridges results back to the CPU runtime via an mpsc::channel + ChannelRecordBatchStream.

IcebergPartitionedTableProvider

• Same optional handle, set via with_io_handle(Handle).
• When set, scan()'s network I/O (load_table + plan_files) is spawned on the IO runtime via Handle::spawn(...).await, preventing DNS/HTTP calls from running on the CPU runtime during DataFusion's physical planning phase.

Why

When DataFusion runs on a dedicated CPU runtime, opendal and iceberg async I/O must not execute on its threads,doing so causes latency spikes and, in some configurations, a panic (nested block_on). This is the same problem solved by dd-datafusion's IOExec wrapper, and this implementation mirrors it directly.

Design notes

• Both types default to io_handle: None; the feature is opt-in via the builder method, no behavior change for existing callers.
• CHANNEL_BUFFER_SIZE = 32 provides backpressure between runtimes (mirrors dd-datafusion's IOExec).
• The JoinHandle from the spawned Parquet-read task is intentionally dropped (detached). Errors are forwarded via the channel; if the task panics, tx is dropped and the consumer sees end-of-stream. This matches dd-datafusion's IOExec behaviour.
• fetch_tasks is extracted as a static method so it can be passed to Handle::spawn without capturing &self.

[notfilippo]

looks very good! great job