parser: columnar Arrow decode for parquet, and --file input

[jameswinegar]

What

Replaces the parquet parser's row-at-a-time record API (RowIter + to_json_value) with parquet::arrow's columnar ParquetRecordBatchReader plus a converter that reproduces the record API's JSON output byte-for-byte. Also finishes wiring the parser's --file seekable-input flag end to end.

Decode (byte-for-byte compatible)

• Columnar decode, one JSON object per row, formatting matched to the record API exactly: decimals and timestamps as strings, binary as base64, non-finite floats as null, nested structs/lists/maps.
• INT96 and INT64-nanosecond timestamps both decode to Arrow Timestamp(Nanosecond) but the record API renders them differently (INT96 → millisecond string, INT64 nanos → raw integer). A per-column plan built from the parquet leaf schema disambiguates; INT96's millisecond value is reconstructed to match Int96::to_millis() across the full nanosecond range, including pre-epoch.
• Arrow types outside a verified allow-list fall back to the record API unchanged (defensive; no type the reader produces today hits it).

Correctness is covered by a test asserting equality with the record API across nulls, nesting, logical types, and all four compression codecs (uncompressed/snappy/gzip/zstd), plus dedicated INT96 (including a pre-epoch value) and nanosecond cases.

--file input

When --file <path> is set, the parser opens the file directly as a seekable input instead of buffering stdin into a temporary file; seekable formats (parquet, Excel) then read footers/row groups without a full pre-read. go/parser gains a ParseFile entry point that passes it. A test asserts --file output matches piped-stdin output across parquet/csv/json/jsonl/avro.

Performance

Ran a local benchmark (not included in this PR) comparing the columnar decode to the record API across column shapes. Preliminary bare-decode results: roughly +12% to +100% depending on shape — largest on string-heavy and few-column tables, smallest on format-bound (timestamp) and very-wide tables. No shape regressed.

Notes

• Adds the arrow feature to the parquet dep, plus arrow-array/-buffer/-schema/half.
• CLI contract: new optional --file flag (default - = stdin, unchanged). The mirrored parser invocation in the connectors repo can pass --file for locally-staged inputs.

Draft — opening for early review.