Optimize REPL read |> count() to use metadata for Parquet/ORC

[aisrael]

When the REPL pipeline is exactly read(path) |> count(), it is now optimized to a single count(path) stage. For Parquet and ORC this uses file metadata only (no row data is read); for Avro and CSV it still streams batches.

Key changes

• optimize_read_then_count() in plan_pipeline_with_state: replaces [Read { path }, Count { path: None }] with [Count { path: Some(path) }].
• read(path) |> select(...) |> count() is unchanged (still three stages); only the two-stage case is optimized.
• Tests: test_plan_pipeline_count_no_auto_print updated for optimized pipeline; test_plan_pipeline_read_select_count_not_optimized added to ensure select-in-between is not optimized.

Made with Cursor