Whole tables in memory

The high-level layer: one parquet_open_table call presents a whole parquet file as a parquet_table, hands its columns back as ordinary Fortran arrays or zero-copy pointers, and reads each column only when something first touches it. A table need not come from a file at all — it can equally be built from scratch in memory and written back out through an ordinary parquet_schema.

• Whole tables in memory: the basics — opening a table and reaching a column with %get or %col, nulls, laziness and what it costs, what a table can tell you about itself, the row and column handles, row selections, string columns, changing a column's type, and the current limitations.
• Opening a table: slices, filters and renaming — reading part of a file (the slice regime), filtering, sorting and qc-checking rows as they are read, renaming columns and carrying units through a read-in MAML, finding which file row a row came from, and what a table can and cannot read.
• Building a table and writing it out — parquet_new_table and %add_column, then parquet_write_table with a schema or without one, its writer options, and carrying the source file's metadata into the output.
• Changing a table — replacing values, editing nulls, adding and dropping columns, filtering/sorting/ranking/appending rows and the detach rule they share, plus searching and ranking a column through a %col pointer.