feat: run the OpenXLA backend on CUDA (GB10) via a source-built IREE runtime (#449)

[inureyes]

Summary

The OpenXLA backend (issue #449) now runs on the GB10 GPU. MLXCEL_BACKEND=xla MLXCEL_XLA_DEVICE=cuda mlxcel generate drives Llama-3.2-1B through IREE on CUDA, token-exact (48/48) vs HF temp-0 at ~5 tok/s (~2.6x the CPU path).

Why a source build

• The prebuilt IREE dist is CPU/Vulkan only: no cuda driver, and its iree-compile has no cuda codegen.
• Vulkan via the dist does not work on the GB10: IREE's Vulkan allocator cannot allocate against NVIDIA's Grace-Blackwell unified memory.
• So CUDA is the GPU path, using a source-built cuda-enabled IREE runtime plus a cuda-capable iree-compile, version-matched. It is a separate, mutually-exclusive build mode keyed on IREE_CUDA_HOME, so the merged CPU (IREE_DIST) path and CI (which build neither) are unchanged.

What changed

• mlxcel-xla/build.rs: cuda mode (IREE_CUDA_HOME) compiles the shim against the source runtime headers with XLA_GATE_CUDA and bakes the cuda iree-compile path (IREE_CUDA_COMPILE, overridable at runtime via MLXCEL_XLA_IREE_COMPILE).
• csrc/xla_iree.c: registers the cuda driver explicitly (guarded by XLA_GATE_CUDA; the unified runtime bundles only the local-task registration).
• root build.rs: the cuda runtime link recipe (the source unified archive already bundles the cuda driver impl, so whole-archive it alone + the registration wrapper + IREE's vendored printf + flatcc; the cuda driver dlopens libcuda, so no link-time -lcuda).
• src/iree.rs: cuda target flags, cuda iree-compile sourcing, and the compiler path in the vmfb cache key.
• runtime.rs: suppress the MLX CPU-fallback footgun warning when the XLA backend is selected (it drives inference through IREE, not MLX, so the message is misleading).
• spike/iree-ffi: the same cuda mode that proved the path token-exact before it was productized.
• README: the source-runtime build recipe.

How to run (local)

# build the cuda-enabled IREE runtime from source (runtime only, no LLVM) — see the crate README
export IREE_CUDA_HOME=/path/to/iree   IREE_CUDA_COMPILE=/path/to/cuda/iree-compile
cargo build --release --features xla-iree
MLXCEL_BACKEND=xla MLXCEL_XLA_DEVICE=cuda ./target/release/mlxcel generate -m <llama-3.2-1b> -p "..." -n 48

Scope

• Still Llama-3.2-1B, prompts capped at the 256-token bucket, greedy, single-sequence (batch-1). Throughput needs batched graphs + a multi-sequence session, a follow-up.
• The cuda runtime build is a local artifact (not committed); the recipe is in the README.

Validation

• dist (CPU) mode still builds (no regression); cuda mode builds, links, and runs end to end on the GB10 at ~5 tok/s.
• default and xla-backend builds, cargo fmt --check, and cargo clippy (default -D warnings + xla-backend) are clean; cargo test --features xla-backend --lib backend:: 6/6.

Refs #449.