[codex] Add v0.6 general-shape plan API

[TruNcat3]

What changed

• add a CUDA-library-style C handle/descriptor/plan API with streams, caller workspace, status codes, logs, exact-key cache, and bounded measurement
• add exact arbitrary FP32/FP64 FFT and compatible uint64 NTT lowering, zero-extended embedding for all operators, and rank-two axis composition
• propagate physical-axis selection into compositions, add direct-output lowering, and select fused logical input for warp-register FP32 FWHT
• add the general-shape benchmark, V100 application profiles, correctness tests, raw measurements, and layered v0.6 documentation

Why

The project architecture treats shape lowering, space-time mapping, physical processing units, and boundary residence as separate design choices. This change extends that model beyond native power-of-two physical plans while preserving explicit semantics and auditable selection.

User and research impact

Applications can use a professional plan workflow for exact, embedded, and rank-two workloads. The selected 32767 -> 32768, batch-256 FWHT composition falls from 0.73772 ms initially to 0.11759 ms after physical-core and boundary selection. Direct exact FFT intentionally abstains to cuFFT at parity rather than presenting Bluestein overhead as the library result.

Structured fused input remains a correct generated candidate but is not enabled by default because it regresses on V100. Cross-GPU transfer and universal superiority remain outside the claim.

Validation

• ctest --output-on-failure: 49/49 passed on V100 with CUDA 11.8
• python3 scripts/check_repository.py
• git diff --check
• five-trial V100 fused-input timing archived in results/general_shape_fused_input_v100_raw.csv