fix(grid): keep wide characters before the cursor when erasing to end of screen

[dylanpulver]

Row::truncate takes a display column, but converted it to a character index with
x - excess_width_until(x). excess_width_until counts the excess of the first x
characters (columns.iter().take(x)), so as soon as more than one wide character precedes the
column it over-counts and the row is cut shorter than asked. position_accounting_for_widechars,
twenty lines below in the same impl, is the converter that gets this right, and
replace_and_pad_end already uses it.

The visible effect is through clear_all_after_cursor (CSI 0 J, which shells emit constantly):
with the cursor anywhere past a second wide character, text before the cursor is deleted.
Measured on main (af38660), first row after the sequence, 20-column grid:

input	expected	on main
日本 then CUP col 3, ESC[0J	日	(empty)
a日本 then CUP col 4, ESC[0J	a日	a
ab then CUP col 3, ESC[0J	ab	ab

Narrow-only rows are unaffected, which is why this has stayed invisible.

Nothing in the suite could see it: every wide-character test in grid_tests.rs
(wide_characters, wide_characters_line_wrap, bash_delete_wide_characters,
fish_wide_characters_override_clock) is an assert_snapshot! of a fixture replay, and the
hand-written ED tests are all ASCII. truncate's formula and those fixtures landed in the same
commit, 44d67de1 (#535, 2021-05-26). Reverting only grid.rs to main fails the one new test
and nothing else; all 1577 lib tests pass with the fix and no snapshot needed regenerating.

Worth saying: the naive version — truncate_position = x, ignoring wide characters — also passes
the new test. It fails five existing ones (delete_char_in_middle_of_line_with_widechar,
insert_character_in_line_with_wide_character, and three search_tests), so the suite pins
over-truncation while being blind to under-truncation. position_accounting_for_widechars is what
satisfies both directions.

Out of scope: the other excess_width_until(x) call, in add_character_at's Ordering::Less arm,
is not affected — there x is at or past the row's display width, so the two readings coincide.
#5240 touches replace_and_pad_end, not truncate; no overlap.

cargo test -p zellij-server --lib (1577 passed), cargo fmt --all --check clean. cargo clippy
at the pinned 1.95.0 already fails on main in output/mod.rs and plugins/unit/plugin_tests.rs;
nothing in the changed files is new.

Found by reading Row's three display-column/character-index converters against each other, not
from normal use. Written with AI assistance (Claude Opus 5, claude-opus-5); every number above
came from running the built test binary.