v0.2.6

Add wgpu-based GPU backend — Metal (macOS), DirectX 12 (Windows), Vulkan (Linux). Add cost model — runtime profiling routes work to CPU or GPU based on data size and transfer cost. Add f32/f64 compute shaders for a broad set of builtins — ones, zeros, rand, reductions, and many more now dispatch to the GPU automatically. Add computation graph pattern scanning — the runtime analyzes your code's computation graph, matches sequences against a library of fusible patterns, and replaces them with optimized GPU kernels automatically. Add 5–6 initial fusion operations including elementwise math chains, where multiple operations collapse into a single GPU kernel eliminating intermediate memory traffic