[ZhuchkaTriplesix]

Closes #1552

PostgreSQL copied every row with row.toList(), SQLite built a list per row, MySQL kept all ResultSetRows and then read each with colAt; CompactResultDataset.fromRawRows then walked the whole list once for the kinds and once per column to pack.

ResultColumnsBuilder (core/database/result_columns_builder.dart) takes the rows as they arrive (addRow, or addCell + endRow for drivers whose rows are not lists) and learns each column's kind on the way. Rules unchanged: all ints int64, all numbers float64, all strings string, all NULL string, else generic; the first 60 non-null cells still go through the driver formatter (#1548). A column can be demoted late (ints that meet a double, anything that meets a string), so ints and doubles are kept in separate buffers: a column that ends up generic still holds its ints as ints, exactly as before. fromRawRows now feeds the builder, so every existing dataset test covers it. MySQL uses a new forEachThenDrain (the drain-after-cap behaviour of takeThenDrain stays, which is built on it).

Tests: kinds and raw values for mixed, late-demoted and all-NULL columns, short rows, cell-by-cell vs whole rows, the formatter case, and 150 random columns checked against the kind rules (some beyond the 256-row first buffer). All of test/core/database, mysql, postgresql, sqlite and workspace tests pass locally.

Measured (JIT test runner, 50k x 20 mixed columns, 3 runs each; rows already in memory as the driver holds them): old path (copy rows, then pack) 124-135 ms, peak RSS over baseline 26-40 MiB; builder 118-121 ms, 21-25 MiB. 5k rows: 16-20 ms vs 17-22 ms, no difference outside noise. So the gain is the intermediate list's memory (about half the peak at 50k) and a small time saving; I did not measure against live servers.