feat: add GPU compute benchmark and GSplat benchmark example

[mvaligursky]

Add a standalone GPU compute benchmark that runs during WebGPU device initialization and a benchmark example for comparing GSplat renderer performance across backends.

Changes:

• Add WebgpuComputeBenchmark class that measures compute throughput (shared memory + barriers + atomics) using raw WebGPU timestamp queries during device creation
• Run benchmark in WebgpuGraphicsDevice.createDevice() and store result as computePerfIndex on the device (milliseconds, or -1 if timestamp queries unavailable)
• Add gaussian-splatting/benchmark example that sequentially tests WebGL2 CPU Sort, WebGPU CPU Sort, WebGPU GPU Sort, and WebGPU Compute renderers
• Example collects CPU frame times, GPU frame times, and per-pass GPU timing breakdowns over 50 measured frames
• Results displayed on screen with GPU info and download button for detailed report including per-pass timings

Performance:

• Benchmark adds ~0.02ms to WebGPU device initialization on Apple M4
• computePerfIndex will be used by GSplat auto-selection (GSPLAT_RENDERER_AUTO) in a follow-up PR to choose between compute and vertex/fragment renderers based on GPU capability