refactor: unify head with tail/sample, add CSV limit, unify column selection

[aisrael]

Aligns head with tail/sample, adds row limit support for CSV, and unifies column selection so CLI and REPL share the same resolution and projection logic.

Commits

1. refactor(head): use build_reader and apply_select_and_display like tail/sample

   • head uses build_reader (with limit = Some(args.number)) instead of read_to_batches.
   • Optional column selection and display via parse_select_step and apply_select_and_display.
   • Explicit format check: only Parquet, Avro, CSV, and ORC supported (same as tail/sample).

2. Enhance CSV and ORC reading capabilities by adding row limit support

   • ReadCsvStep gains an optional limit; build_reader passes limit for CSV so head can limit rows without loading the full file.
   • head sets offset 0 when limiting ORC for correct row selection.
   • select_columns_to_batches (REPL) now uses the same resolution and Arrow project as the streaming SelectColumnsStep (no DataFusion in this path); REPL exec_select calls it synchronously.

Summary

Area	Change
head	Same reader → select → display pipeline as tail/sample
CSV	build_reader + ReadCsvStep support limit for streaming-style head
Column selection	One resolution path (resolve_column_specs) and one projection method (Arrow project) for both CLI and REPL
REPL select	Sync select_columns_to_batches; empty specs return batches unchanged