Sync with longbridge/gpui-kit 21113cae, gpui-fast 4afc87b

[trancong12102]

Takes longbridge/gpui-kit 201b55a4..21113cae and moves the gpui-fast pin to 4afc87b. This fork carries its commits rebased on upstream, which is also how cargo xtask upstream sync lands it, so this branch is the 33 fork commits rebased onto 21113cae, plus one new commit. Land it by moving main to the branch head (--force-with-lease on b06a0953), not with a merge commit. A merge would bring every fork commit in twice.

Upstream commits taken

• shell: update QuickJS JIT to 0.12.11 longbridge/gpui-kit#3338, QuickJS JIT 0.12.11 for gpui-shell. Lockfile and manifests only. Upstream measures it as faster on every workload than 0.12.9. 0.12.10 is skipped because it regressed. Fine to take as is.
• docs: Document macOS font-kit requirement longbridge/gpui-kit#3339, docs on macOS font-kit. English and Chinese. Docs only.
• text: Avoid remeasuring scroll-table column widths longbridge/gpui-kit#3318, column-width cache for scroll-layout Markdown/HTML tables. Widths are cached on the parsed table, keyed on the window's text system and typography, and only custom cells are measured again. In retained mode a table view that nothing changed is not built again at all, so the cache helps when the view is built again (a scroll or an edit nearby). It costs about 0.5 KB per table. One gap: it doesn't notice TextSystem::add_fonts, because GPUI exposes no font generation to read. Same tradeoff upstream accepted; taken.
• markdown: add source & range_for_source longbridge/gpui-kit#3281, RenderedText::source / range_for_source. The source map lives in RenderedIndex, which is built lazily (get_or_init) only when someone asks for the rendered text. The parse itself only records one more flag per segment, so streamed Markdown pays nothing extra.
• text_view: Let fenced code blocks scroll inside a height cap longbridge/gpui-kit#3322, fenced code blocks scroll inside a height cap. Opt-in via overflow.y = Scroll on the code_block refinement. Useful for Slopty's conversation face.

Fork commits changed by the rebase

• perf(text): parse only the last block when set_text extends Markdown renamed MAX_SYNC_FULL_REPLACE_BYTES to MAX_SYNC_PARSE_BYTES, and markdown: add source & range_for_source longbridge/gpui-kit#3281's new tests used the old name. It also parses a short append's tail at once, where upstream parses it in the background. So markdown: add source & range_for_source longbridge/gpui-kit#3281's a_snapshot_converts_against_its_own_source now expects the new source to be rendered straight after push_str. Its other checks are unchanged: the earlier snapshot still converts against its own source. Both fixes are folded into that commit.
• New: build: take gpui-fast 4afc87b. The fork's sync with longbridge/gpui-fast 92a9b0f (Update to use better implementation of selection support for Input. longbridge/gpui-kit#31-Failed to build on Linux longbridge/gpui-kit#34) and zed f8c2cc84 (#64990).
• Every other commit is identical (git range-diff).

Verification (macOS, at d6e51664)

• CI's test command, cargo test --workspace --exclude gpui-shell --features gpui-component-story/test-support --locked (example crates excluded as CI does): 2602 passed.
• --test rendering: 10 passed (Metal).
• cargo clippy --workspace --exclude gpui-shell --locked -- --deny warnings: clean.
• cargo xtask upstream sync --only gpui-kit --no-push (Slopty): passes its workspace clippy (--all-targets), the story + recipes tests and the shell tests.

Known flake, already there before this sync

text::window_selection::tests::drag_inside_dialog_still_selects_its_text fails now and then:

• 1/400 runs at the old gpui-fast pin ca9be59, and 0-9 per 200 at 4afc87b, depending on machine load.
• Likely cause: GPUI's animation element times itself with Instant::now(), but the test's settle only advances the test executor's clock. The dialog's opening animation can therefore still be moving when the drag lands.

Co-Authored-By: Claude Code