fix(render): disable browser scroll anchoring on the table holder - #4935

Problem

Chrome's scroll anchoring competes with the virtual renderer's padding writes. At the
bottom of a long list it pins scrollTop and the two fight in an infinite
oscillation — the renderer adjusts padding, anchoring restores the scroll offset, and
the cycle repeats every frame.

Fix

Set overflow-anchor: none on the table holder. The virtual renderer positions content
itself, so browser anchoring has nothing useful to contribute.

Test

test/e2e asserts the table holder disables scroll anchoring.

This PR also corrects the scroll-jump e2e measurement, because with anchoring off the
existing measurement no longer works and, once corrected, it reports a real defect on
current master. Two changes, in the last two commits:

1. The grouped case stepped 250px in a 275px viewport. It tracks one row across each
   step, so that row must be in the DOM before and after. On master anchoring cancels most
   of every step — the test asks for 250px and gets 46, 14, 38, 54 — so the row survives by
   accident. With anchoring off the scroll does what it was asked, the whole window turns
   over, and no row is present in both states, so nothing is measured. The step is now
   80px, which keeps a trackable row.
2. The jump was measured as moved − scrolled, where scrolled is the holder's own
   scrollTop delta. A virtual renderer revises its total scrollHeight as it measures
   real row heights, and scrollTop is read from the top of that changing document, so it
   shifts by the height correction while nothing on screen moves. It now compares the
   row's screen movement against the requested scroll distance, which has no such term.

Both tolerances, 20px and 5px, are unchanged. Verified against three other states: the new
metric still fails at 280px on the code before the #3654 fix, so it keeps that coverage; it
passes on this branch; and it fails at 124px on current master, where anchoring silently
cancels part of each requested scroll. That last figure is independent evidence for this
fix — the defect was previously invisible because the measurement absorbed it.

Performance

Neutral. 500k rows, K=5, medians: