gpui: count primitives into draw order instead of comparing them

[sunli829]

What

Scene::finish puts every kind of primitive in drawing order. It sorted an
index vector with sort_unstable_by_key, reading each primitive's key out of
the 112–168-byte primitive on every comparison, then gathered.

Draw orders are few and small (one past the greatest among the primitives a
primitive overlaps, or its layer's): a frame of workspace-quotes has 3324
glyphs over 21 orders. So, in fast/scene.rs only:

• Primitives other than sprites are counted into their order's run, one
  pass to count and one to place, and left alone when already in order.
  Orders spread more than four times wider than the primitives are numerous
  are sorted as packed u64 keys instead, so the counts stay short.
• Sprites sort one packed u128 of order, texture, tile and index, and
  are left alone when already in order.

Both give exactly the order the old sort did, so batches and drawing are
unchanged. No upstream file changes.

Correctness

• cargo test -p gpui --features test-support --lib: 646 passed. New tests
  compare both sorts with a reference stable sort over random scenes (empty,
  one item, in order, counted, spread too wide), check which way each case
  was sorted and the threshold at its edge, and that the packed sprite key
  keeps every bit of order, texture and tile.
• gpui_perf --headless --verify: all 55 checks identical.
• script/check-upstream: 0 violations.

Performance

The sort on real keys dumped from frames (single core, hot cache):

Keys	Before	After
Real window, glyphs (1871)	37.8 µs	~20 µs
Real window, quads (491)	9.1 µs	3.9 µs

Headless, main-thread instructions per frame, retained views on, all 42
scenarios: every one lower, median −2.3%.

Scenario	Change
list-filter	−17%
settings-toast / form-hover	−11% / −10%
list-variable-scroll / list-uniform-scroll	−9%
chat-scroll	−4.6%
workspace-*	−1.9% to −3.0%
layout-*, table-ticks-many	−0.1% to −0.6%

Real window (--auto --only WorkspaceQuotes, 3 alternating runs of 600
frames): instructions p50 16.5M → 16.1M with retention on, 22.7M → 22.3M
off. Wall-clock CPU time on this machine is noisier than the difference.

🤖 Generated with Claude Code