feat(gpu): add bounded high-precision export rendering

[VailElla]

Description

Add a bounded, crate-private high-precision GPU render path for the follow-up JPEG XL export work. The normal preview and export path keeps its existing 8-bit contract, while the new path renders tile output as rgba16float and returns RGB/RGBA 32-bit float image data without an 8-bit intermediate.

This draft also bounds reusable compute textures to one overlapped processing tile instead of allocating six full-image textures for large inputs.

Type of Change

• Bug fix
• New feature
• Breaking change
• Performance improvement
• Code refactoring
• Documentation update
• UI/UX improvement
• Build/CI or Dependency update

Changes Made

• Added a high-precision compute pipeline variant with an rgba16float storage target while deferring integer dithering to the follow-up encoder.
• Generalized tiled GPU readback for both 4-byte RGBA8 and 8-byte RGBA16F pixels.
• Converted RGBA16F readback into RGB32F or RGBA32F while preserving the source alpha contract.
• Kept the high-precision entry crate-private until the follow-up JXL backend wires it into export.
• Bounded reusable ping-pong, blur, and tile-output textures to a 2304 × 2304 overlapped tile; full-size display textures remain unchanged.
• Avoided a full RGBA32F intermediate when the input is already RGB32F or RGBA32F.
• Added three unit tests for high-precision shader behavior, RGB32F upload conversion, and the bounded tile extent.

Screenshots/Videos

N/A — this is a backend-only foundation and does not change the current UI or exported output by itself.

Testing

• I have tested these changes locally and confirmed that they work as expected without issues

Test Configuration:

• OS: macOS 26.5.2
• Hardware: Apple M4 Pro, 14 logical cores, 48 GiB RAM

Automated:

• cargo +1.96.1 fmt --manifest-path src-tauri/Cargo.toml --check — passed
• cargo +1.96.1 check --manifest-path src-tauri/Cargo.toml — passed; warning set matches current main with no new warnings
• cargo +1.96.1 test --manifest-path src-tauri/Cargo.toml -- --nocapture — 3 passed, 0 failed
• cargo +1.96.1 test --manifest-path src-tauri/Cargo.toml gpu_processing:: -- --nocapture — 3 passed, 0 failed after final cleanup
• cargo +1.96.1 clippy --manifest-path src-tauri/Cargo.toml --all-targets --all-features — passed; only current macOS baseline warnings remain
• npm ci — completed with 0 vulnerabilities
• npm run build — passed
• git diff --check — passed

Checklist

• My code follows the project's code style
• I haven't added unnecessary AI-generated code comments
• My changes generate no new warnings or errors

Additional Notes

• Based directly on main at 41d92c8f and kept to two focused commits touching only src-tauri/src/gpu_processing.rs.
• This is an internal prerequisite for the smaller JXL export series related to FEATURE: Export JPEG XL & WepB with EXIF metadata #1322 and BUG: Exporting JPEG XL causes massive memory consumption #1342; it does not close either issue by itself.
• The crate-private high-precision entry is intentionally not wired into export in this draft. JXL encoding, metadata, export controls, and end-to-end real-image validation remain in follow-up PRs.
• The existing 8-bit preview/display path remains the default and retains its current output format.

AI Disclaimer:

Please state the involvement of AI in this PR:

• This PR is entirely AI-generated
• This PR is AI-generated but guided by a human
• This PR was handwritten with AI assistance (spell check, logic suggestions, error resolving)
• This PR contains only blood, sweat, and coffee (AI-free)

[VailElla]

Refresh validation (2026-07-25): rebased onto current upstream main and force-with-lease updated. Added an adapter capability gate for the RGBA16Float high-precision path.

Compatibility behavior:

• adapters whose RGBA16Float format supports texture binding, storage binding, and copy-source create the high-precision pipeline;
• other adapters keep the existing RGBA8 preview/export pipeline, while a request for the internal high-precision path returns a specific unsupported-GPU error instead of preventing normal GPU initialization.

Memory regression coverage now models a 100 MP (11,600 x 8,600) input: five reusable RGBA16F intermediates remain capped at 2,304 x 2,304 (<220 MiB combined), versus >3 GiB if those intermediates were full-image allocations. Full-size RGBA8 display textures are intentionally unchanged.

cargo test gpu_processing:: --lib passed: 4 passed, 0 failed.