Closes #623.

Summary

This adds a stable native-read descriptor for ordinary Lance Spark scans so native engines can consume Spark-planned Lance reads without depending on Lance Spark internals.

The descriptor captures:

• Dataset URI and resolved dataset version.
• Spark read schema JSON and projected read schema JSON.
• Projected columns, pushed filter SQL, limit/offset, batch size, and storage options.
• Per-partition native splits with Lance fragment IDs.
• Explicit fallback reasons when a scan cannot be represented by the minimal v1 descriptor.

The v1 scope is ordinary table reads only. Search/hybrid search, index-backed execution descriptors, aggregation pushdown, metadata/version columns, and namespace-backed credential refresh remain fallback/future work.

Why LanceScan carries the descriptor state

This PR adds several parameters to the LanceScan constructor because LanceScan is the object Spark keeps inside BatchScanExec after planning. A native consumer such as Comet sees that final BatchScanExec(scan = LanceScan) object, not the earlier LanceScanBuilder or catalog planning context. Therefore nativeScanPlan() needs the complete, already-resolved scan snapshot on LanceScan itself.

The goal is not to make the constructor a broad public API. The goal is to avoid asking native consumers to infer or recompute Lance Spark planning semantics from partial state. Re-planning later would be risky because it could reopen a different dataset version, use different storage options, produce different fragments, or miss fallback-only state such as pushed TopN/aggregation.

The added state is the minimum needed to describe or reject the native v1 scan accurately:

• sparkReadSchema and schema keep the Spark-visible schema and projected read schema separate, which matters when Spark-facing fields differ from the physical/projection schema.
• readOptions provides dataset URI, resolved version, batch size, table/catalog identifiers, and user storage options. The resolved version is required so native execution cannot drift to a newer Lance snapshot.
• whereConditions, limit, and offset are serialized into the descriptor when v1 supports them.
• topNSortOrders and pushedAggregation are carried even though v1 falls back for them, because the descriptor must reject those scans explicitly instead of silently dropping semantics.
• pushedPredicates, zonemap stats, surviving fragment IDs, precomputed splits, and fragment row counts preserve the fragment-pruning and limit-pruning decisions Lance Spark already made on the driver.
• activeShardingExpression and fragmentShardingKeys preserve the existing partitioning/reporting contract used by Lance Spark planning.
• initialStorageOptions, namespaceImpl, and namespaceProperties preserve storage option precedence and the namespace context that workers/native readers need for the same dataset access path.

A follow-up cleanup could wrap this constructor state into an internal immutable scan-state object if reviewers prefer that shape. This PR keeps the change direct so the descriptor contract and tests are easy to review first.

Testing

• ./mvnw -pl lance-spark-base_2.12 -Dtest=LanceScanTest -Dspotless.skip=true test
• ./mvnw -pl lance-spark-4.1_2.13 -am -Dtest=LanceScanTest -Dspotless.skip=true -Dsurefire.failIfNoSpecifiedTests=false test
• ./mvnw -pl lance-spark-base_2.12 spotless:check