baracuda-transformer-engine vs baracuda-transformer-engine-sys vs atomr-accel

Safe Rust wrapper for baracuda's port of NVIDIA TransformerEngine's FP8 cast/transpose + delayed-scaling recipe primitives. Provides `Fp8Recipe` (delayed-scaling state with amax history), `Fp8CastPlan` for {f32, f16, bf16} → FP8 with running amax, `Fp8DequantPlan` for FP8 → {f32, f16, bf16}. Cast/recipe subset only — `normalization` / `fused_rope` / `fused_attn` / `fused_softmax` / `activation` / `gemm` skipped (overlap existing baracuda phases). NO cuDNN dep, NO pybind11. On Ada (sm_89) the FP8 wins are bandwidth-saving only (KV cache, weights); FP8 tensor-core math throughput equals BF16. Forward-compatible with Hopper / Blackwell where the compute wins also materialize.

Build + raw FFI bindings to baracuda's port of NVIDIA TransformerEngine's FP8 cast/transpose + delayed-scaling recipe primitives. Cast/recipe subset only — `normalization` / `fused_rope` / `fused_attn` / `fused_softmax` / `activation` / `gemm` deliberately skipped (overlap existing baracuda Phase 3/5/14/17/30/31/36/41/42). NO cuDNN dep (recipe + cast paths don't need it; `fused_attn` would, and we skip it); NO pybind11 (the safe wrapper lives in `baracuda-transformer-engine` and exposes a raw C ABI defined in `csrc/baracuda_te_shim.cu`). Apache-2.0 per upstream — see `ATTRIBUTION.md`.

Backend-agnostic compute-acceleration core. Defines the AccelBackend trait, AccelRef<T> typed pointers, AccelError enum, and CompletionStrategy — the abstraction layer that lets atomr-accel-cuda (NVIDIA), and future ROCm / Metal / oneAPI / Vulkan backends plug into the same actor surface.