Add STDE deferred items: const-generic diagonal, dense STDE, GPU Taylor

[gvonness-apolitical]

Summary

• Add diagonal_kth_order_const<ORDER> — stack-allocated const-generic variant of diagonal_kth_order for compile-time-known derivative order, avoiding TaylorDyn arena overhead
• Add dense_stde_2nd — dense STDE for positive-definite 2nd-order operators using Cholesky-transformed Gaussian directions (tr(C · H) estimation)
• Add GPU-accelerated 2nd-order Taylor forward propagation via new WGSL shader (taylor_forward_2nd.wgsl) and CUDA kernel (taylor_eval.cu), propagating (c0, c1, c2) triples through the tape
• Integrate taylor_forward_2nd_batch into both WgpuContext and CudaContext
• Add high-level GPU STDE wrappers: laplacian_gpu, hessian_diagonal_gpu, laplacian_with_control_gpu
• Add comprehensive tests: const-generic diagonal, dense STDE, per-opcode GPU Taylor verification, end-to-end GPU STDE cross-validation against CPU
• Add benchmarks for const-generic vs dynamic diagonal and GPU vs CPU STDE
• Update documentation across algorithms.md, README.md, and module docs

Test plan

• cargo test --features "stde,diffop,gpu-wgpu" — 640 tests pass, 0 failures
• cargo build --features "stde,diffop,gpu-wgpu" — clean build
• Const-generic diagonal matches dynamic variant for k=2,3,4
• Dense STDE with L=I matches Hutchinson Laplacian
• GPU Taylor c0/c1/c2 match CPU taylor_jet_2nd for all opcode categories
• GPU laplacian_gpu matches CPU laplacian (tolerance ~1e-4)
• GPU hessian_diagonal_gpu matches CPU hessian_diagonal (tolerance ~1e-4)