perf(desktop-gpui): fast row blit for opaque picker thumbnails

[nikolas-sapa]

Description

In fit_centered, centering a fitted image onto a newly created transparent canvas currently uses image::imageops::overlay.

Because canvas is already empty ([0, 0, 0, 0]), source-over alpha blending onto it is an identity operation:

• For opaque source pixels (alpha == 255), overlay runs floating-point alpha calculations, bounds checks, and put_pixel per-pixel.
• By detecting if the source image is opaque using the existing SIMD is_opaque helper, we can copy rows directly via copy_from_slice, completely bypassing the per-pixel float math and bounds checks.
• When transparency is present, it falls back to imageops::overlay to preserve identical behavior.

Benchmarks

Tested on Apple Silicon using cargo bench --bench thumbnail_fit:

Workload	Width x Height	Baseline	Candidate	Speedup
fitted-wide	320x134	65.1 µs	8.2 µs	~7.9x
fitted-tall	218x180	73.4 µs	11.6 µs	~6.3x
fitted-square	180x180	47.4 µs	8.3 µs	~5.7x
fitted-near-full	320x175	77.6 µs	9.3 µs	~8.3x
upscale	160x90	753 µs	683 µs	1.10x
downscale	640x360	1780 µs	1695 µs	1.05x
downscale-1080p	1920x1080	10.8 ms	10.4 ms	1.04x

Verification

• cargo test -p cap-gpui-kernels: all 30 tests pass (including fitted_sources_keep_exact_centering_and_transparent_padding, resized_sources_match_lanczos_reference, exact_size_preserves_hidden_rgb_and_source_pixels).
• Exact bit-level parity verified across all workloads.

Confidence Score: 5/5

This PR appears safe to merge.

Summary

fit_centered now copies opaque thumbnail rows directly into the transparent canvas.

• Fitted opaque thumbnails now copy rows into the canvas.

_{Reviews (2) · Last reviewed commit: "Merge branch 'main' into perf/gpui-fast-..." · Reviewed by Greptile}

[vercel]

@nikolas-sapa is attempting to deploy a commit to the Cap Software Inc Team on Vercel.

A member of the Team first needs to authorize it.

[richiemcilroy]

@greptileai review