Implement ROADMAP.md Phase 4: all 5 deferred features

[gvonness-apolitical]

Summary

• 4.1 Indefinite dense STDE: dense_stde_2nd_indefinite() — eigendecomposition + sign-splitting for arbitrary symmetric C matrices, with epsilon clamping and single-pass optimization for same-sign eigenvalues (requires stde + nalgebra)
• 4.5 GpuBackend trait method: taylor_forward_2nd_batch lifted into GpuBackend trait; all stde_gpu functions now generic over B: GpuBackend; old backend-specific functions deprecated
• 4.4 Generic laplacian_with_control_gpu: works with any GpuBackend, enabling CUDA variance-reduced Laplacian
• 4.3 Chunked GPU Taylor dispatch: taylor_forward_2nd_batch_chunked — splits large batches by buffer size (128 MiB default) and WebGPU workgroup dispatch limits (65535×256)
• 4.2 General-K GPU Taylor kernels: runtime codegen (taylor_codegen.rs) generates K-specialized WGSL/CUDA shaders for K=1..5 with fully unrolled Cauchy products and recurrences for all 43 opcodes; taylor_forward_kth_batch returns TaylorKthBatchResult

Test plan

• 6 indefinite STDE tests (PD cross-validation, diagonal indefinite, full indefinite, all-negative, zero matrix, near-zero eigenvalue clamping)
• 19 existing GPU STDE tests pass (trait method backward compat)
• 4 chunked dispatch tests (single-chunk, multi-chunk with c1s, exact boundary, zero batch)
• 5 general-K Taylor tests (polynomial all orders K=1..5, K=3 cross-validation, exp higher-order, multi-batch deinterleaving, unsupported K error)
• CUDA tested on vast.ai A100 (CUDA 12.8) — compilation, trait impl, doctest all pass
• Full test suite: cargo test --features "bytecode,taylor,laurent,stde,serde,faer,nalgebra,ndarray,parallel,diffop,gpu-wgpu" — all pass
• cargo clippy -- -D warnings clean
• cargo fmt --check clean