Avoid empty-mode extent shortcut in sampling

Copilot review pointed out that dataZoom empty mode writes the zoom window into processed data approximate extents. Using that processed extent to skip the visible-count scan can therefore keep the sampling rate tied to the full dataset.

Use processed extents only when the data count proves the series was actually filtered. For none, empty, and fully visible data, use raw data extents to decide whether the count scan is necessary.

Constraint: dataZoom filterMode empty preserves item count while replacing out-of-window axis values with NaN
Rejected: Always trust processed approximate extent | empty mode intentionally overwrites it with the zoom window
Confidence: high
Scope-risk: narrow
Directive: Do not use processed approximate extent as a fully-visible shortcut when count still matches raw count
Tested: ./node_modules/.bin/eslint --quiet src/processor/dataSample.ts; ./node_modules/.bin/tsc --noEmit --pretty false; git diff --check -- src/processor/dataSample.ts; Chrome headless HTML regression screenshot
Not-tested: Full visual regression suite