feat: CometNativeScan per-partition plan serde

[mbutrovich]

Which issue does this PR close?

Partially address #3510. Subset of what #3446 was.

Rationale for this change

CometNativeScan currently serializes all file partitions upfront in the driver and sends the entire list to every executor task, which then indexes by partition ID. This approach has several drawbacks:

1. Memory overhead: For scans with many partitions, serializing all files for all partitions consumes significant driver memory
2. Network waste: Each executor task receives data for all partitions but only uses its own partition's files
3. Inconsistent with Iceberg: PR feat: CometExecRDD supports per-partition plan data, reduce Iceberg native scan serialization, add DPP [iceberg]  #3349 introduced split-mode serialization for Iceberg scans, but NativeScan still uses the old pattern

This PR extends the PlanDataInjector framework to CometNativeScan, bringing it in line with the Iceberg implementation.

What changes are included in this PR?

• Split NativeScan serialization into common data (schemas, filters, projections) and per-partition data (file lists)
• Common data serialized once at planning time and embedded in the native operator
• Per-partition data lazily serialized at execution time via @transient lazy val
• Each executor task receives only its partition's files via runtime injection

How are these changes tested?

Existing tests.

[andygrove]

LGTM. Thanks @mbutrovich!