New log UI - #6939

The log viewer that follows remounts its virtualiser whenever a fresh size
cache arrives. 0.48.6 starts observing from a `tick()` microtask with no guard,
so a virtualiser destroyed before that microtask resolves still attaches its
observers to a detached container; 0.50.4 carries the fix (inokawa/virtua#914).

Bumped in every workspace that depends on virtua so the monorepo resolves one
copy. The API surface used here is unchanged between the two versions.

`useReverseScrollContainer` needed a fixed number of scroll passes to reach the
bottom, which is why the scroll-to-bottom button had to be pressed twice: the
target is `scrollHeight`, scrolling there mounts more content, and that moves
the target. It also read user intent off scroll geometry, so a virtualiser
correcting its own `scrollTop` after measuring a row cancelled the scroll that
was doing the correcting.

`useStickToBottom` settles by iterating to a fixed point instead, and takes
release exclusively from input events, which carry their author by construction.
Two collaborators keep the rest out of the anchor itself:

  releaseGestures    what counts as the user taking over (wheel, keys, touch)
  viewportDormancy   when the container's geometry may be believed at all

Dormancy is the tab-switch case: a hidden container keeps receiving content
against a frozen viewport, and the catch-up scroll event on return used to read
as "the user scrolled up" and detach a streaming log for good.

Both anchors are exported while the migration runs. `TabAdHocQuery` moves over
here because it drives no FAB; the rest share `ScrollDownFab`, whose contract
changes with the log viewer.

…iewport

The rendered window was `visibleCount - 1` rows, added as a workaround for what
was recorded as a flickering issue when scrolling to the bottom of a long list.

The flicker was the missing row. `indexOffset` can start half a row above the
viewport and the count can round half a row down; `visibleCount` carries two
spare rows to pay for both, so dropping one leaves a strip up to a row tall
blank along the bottom edge at some offsets — which reads as the list failing
to keep up rather than as arithmetic.

The spec sweeps two full row heights of scroll offsets, since a window one row
short only leaves a gap at some offsets within the rounding period.

The decoder emitted each line with its `\n` or `\r\n` still attached, and every
consumer then had to undo it — the log host was passing an empty join separator
to the copy builder precisely to avoid doubling them.

A line that carries its own newline is not a line. Worse for the log viewer that
follows: its height model predicts wrapped row heights over printable ASCII and
tabs, and declines anything else. A surviving terminator puts every row outside
that grid, so the model declines all of them and the size cache is empty for the
whole stream. A stray `\r` is the nastier half — invisible in the rendered row,
identical in effect.

The `\r` test is bounded by `lineStart`, so an empty line, whose whole record is
the terminator, cannot read the previous line's last byte.

`LogList` virtualised by line and could not keep a row mounted, which cost it
three separate behaviours: a selection collapsed the moment either endpoint
scrolled out, a copy silently truncated to whatever happened to be mounted, and
a search highlight had to be re-attempted every frame until the scroll happened
to mount its row. Row heights depend on the container width, and a virtualiser
only measures what it mounts, so halving the width left the reported scroll
range 42.8% short on a 600-line corpus of mixed-length lines — two fifths of
the log unreachable for the rest of the session.

LogView renders in 50-line chunks and seeds the virtualiser with a full size
cache instead of letting it discover heights. Monospace rows make wrapped
heights arithmetic, so the offsets are right in the rebuild's first frame; that
is what lets a reading position be restored with one assignment and a search
jump land without a convergence loop. The parts are separated by concern:

  logChunks       cutting on absolute-line boundaries, so keys survive eviction
  logRowMetrics   the wrap model and the predicted layout
  logLayout       when to re-predict, and restoring the view after the rebuild
  logLineAlign    converging a search jump on the centre of the viewport
  selectionPin    holding the chunks a live selection touches
  logCopy         reconstructing clipboard text from the source lines
  selectScope     replaces 682 lines of virtualSelect with native selection

Adopting the anchor everywhere finishes the migration started two commits back:
`ScrollDownFab` now reads the raw `stuck` value, so it disappears on the press
that re-arms rather than a frame later, and `Query` and `ChangeStream` move with
it. `useReverseScrollContainer`, `userSelect` and the already-orphaned
`ReverseScrollList` go with their last consumers.

Verified against the running console on a 6000-line pipeline log as well as the
suites: streaming stays anchored, one press reaches the true end, and a width
change never paints a frame away from it.