test: close the debt of #1162, #1164 and #1165

[ZhuchkaTriplesix]

Closes #1162
Closes #1164
Refs #1165

Changes

• ui(results): UI kit in results toolbar, charts and groupings (segmented control, icon buttons, export, empty states) #1162: the guard checked only the top-level results files. Two files that the results tab pulls in, data_grid_filter_bar.dart and data_grid_value_panel.dart, still used Material icon buttons. They now use QueryaIconButton (dense), and the guard scans them. The MQL console (mongo_query_workspace.dart) was already in the guard.
• feat(explain): visual query plan (Plan | Text) for PostgreSQL, MySQL, SQLite #1164: parser tests on recorded output for a three-table MySQL join (tree, per-table cost, highlighted table, estimates, equal costs, no query block) and a three-step SQLite plan (tree, detail text, no highlight). PostgreSQL was already covered.
• feat(stats): sparklines for polled metrics on Overview dashboards #1165: the stats history recorder moved out of the view into recordServerStats, which is tested for the first and second poll, a restart, missing databases and missing counts. Sparkline widget tests cover the tooltip before the first sample, the newest value, whole numbers and a flat series.

Not covered (#1165 stays open)

• "No extra queries per poll" needs a counting fake service. PostgresService is a singleton with no seam, so that test is not written.
• Polling stopping when Overview is hidden is covered by ticker_gated_polling_test.dart, not by a test of this view.

[github-actions]

Grid scroll benchmark

metric	base	PR	change
p50	42.20 ms	42.37 ms	+0.4%
p90	51.28 ms	51.15 ms	-0.3%
p99	63.82 ms	69.07 ms	+8.2% ⚠️
stutters	223.00	222.00	-0.4%

Informational only (threshold 5%). Shared CI runners are noisy; re-run before trusting a single result.

[github-actions]

Grid scroll benchmark

metric	base	PR	change
p50	42.38 ms	42.18 ms	-0.5%
p90	50.56 ms	50.94 ms	+0.8%
p99	61.64 ms	65.58 ms	+6.4% ⚠️
stutters	230.00	226.00	-1.7%

Informational only (threshold 5%). Shared CI runners are noisy; re-run before trusting a single result.