[ZhuchkaTriplesix]

Closes #1549

The issue asked to measure first. benchmark/result_consumers_bench.dart (flutter test benchmark/result_consumers_bench.dart, headless) runs sort, filter and CSV export over a 50,000 x 20 CompactResultDataset (6 int, 4 double, 10 string columns) once through asLazyRowList() and once through a List<List<String>> built from it. Median of 5 after a warm-up, two runs:

consumer	materialized	lazy
sort, int column	24-26 ms	24-27 ms
sort, double column	23-25 ms	25-30 ms
sort, string column	181-182 ms	162-171 ms
filter c0 > 5000	6-7 ms	2-6 ms
filter, free text	45-50 ms	85-90 ms
CSV export	102-112 ms	132-133 ms

Decision: keep List<List<String>> at the consumer boundary. The lazy rows cost nothing for sort and typed filters, and 25-40 ms more for a free-text filter and the export, which run once per user action. Moving them to typed accessors would rewrite the sorter, the filter engine and the exporters for that, and the app already holds only the columnar dataset (the 68.6 -> 17.5 MiB figure of the issue is what a run keeps).

Scope item 3, where the dataset is copied to strings today:

• DataGridStagingBuffer copies rows.map(List.unmodifiable), but it is built for table-browser pages and table edits, not for SQL result sets (those stay lazy).
• Sorting returns lazy row wrappers (one small object per row), not strings.
• The isolate hand-off of a sort or filter sends the dataset's typed lists, not strings; checked with a dataset that has a generic column with a formatter closure (it sorts in the isolate).

Memory is not measured by this bench (RSS in the test runner is too noisy); the issue's AOT numbers stand.