Defer parquet scan column planning

[platypii]

Summary

parquetScan is a lazy API, but it previously built column read plans eagerly for every candidate row group.

This refactors planning into two stages:

• parquetPlanGroups selects candidate row groups and row ranges.
• parquetPlanGroup builds column chunk and byte-range plans for one group.
• parquetPlan composes both stages, preserving existing parquetRead behavior.
• parquetScan now performs only group planning during initialization and creates a one-column plan when readColumn is called.

This avoids inspecting and planning unused columns while keeping row-group pruning, page ranges, offset indexes, and scan caching behavior unchanged.

Performance

Local benchmark using a 170 MB Parquet file with 129 row groups:

Case	master	This PR
Scan initialization	0.0334 ms	0.0101 ms
Initialize + read first range	0.375 ms	0.341 ms
Read one column across all groups	33.28 ms	34.14 ms

Scan initialization is approximately 3.3× faster. Full-column read performance is effectively unchanged within benchmark noise.

The existing full-file parquetReadObjects benchmark improved from a median of 1,111 ms to 1,025 ms, though it does not exercise the parquetScan path directly.