Experimental GPU-driven rendering

[AzurIce]

Closes #139

Summary

Merge all VItem data into a single set of contiguous GPU buffers and use instanced drawing to render all VItems in one draw call, eliminating per-VItem CPU submission overhead.

Approach

• Each frame, pack all VItem points, colors, and widths into contiguous buffers with an ItemInfo index table
• Compute shader binary-searches item_infos to find each point's owning item, performs 3D→2D projection + atomic clip box updates
• Render passes use draw(0..4, 0..N) instanced drawing; vertex shader looks up per-item data via instance_index

New files

• primitives/merged_vitem.rs — MergedVItemBuffer: CPU-side data packing, bind group management
• pipelines/merged_vitem.rs — compute / depth / color pipeline definitions
• shaders/merged_vitem_compute.wgsl — merged compute shader (binary search + atomic clip box)
• shaders/merged_vitem.wgsl — merged render shader (instanced vertex + SDF fragment)
• Renderer::render_store_merged() — merged rendering entry point that bypasses the render graph

Benchmark comparison

The original rendering path is kept intact for A/B comparison.

CPU submission time (the bottleneck)

VItem count	Original (per-VItem)	Merged	Speedup
25	1.61 ms	1.64 ms	~1x
100	4.14 ms	1.70 ms	2.4x
400	25.2 ms	1.79 ms	14x
1600	92.7 ms	1.86 ms	50x
3600	220 ms	1.90 ms	116x

Total time (CPU + GPU)

VItem count	Original	Merged	Speedup
25	5.6 ms	3.8 ms	1.5x
100	8.9 ms	6.5 ms	1.4x
400	22.3 ms	5.9 ms	3.8x
1600	88.8 ms	4.6 ms	19x
3600	256 ms	5.0 ms	51x

CPU submission time is now essentially constant (~1.9ms) regardless of VItem count. At 3600 VItems, total frame time drops from 256ms to 5ms.

Visual correctness

Both paths produce identical output for the same scene (including OIT transparency and depth ordering).