Can we avoid JVM round-trips in some cases for `native_iceberg_compat` scans?

[andygrove]

What is the problem the feature request solves?

native_iceberg_compat reads parquet in Rust via NativeBatchReader, then imports the data to JVM via Arrow FFI, then re-exports it back to native for execution by the next operator in the plan. Although Arrow FFI is zero-copy for the data, there is still serialization overhead for the schema.

I am wondering if this round-trip can be avoided in some cases

The following profiling information is from running async-profiler.

Describe the potential solution

No response

Additional context

No response