[inureyes]

Summary

Part of the OpenXLA low-precision performance epic (#570). Wraps the source-built IREE runtime setup (scripts/iree/setup-cuda.sh, scripts/iree/setup-macos.sh) in idempotent make targets so the measurement environment for the rest of the epic is reproducible from a fresh checkout rather than a local uncommitted artifact.

What changed

• Makefile targets. make iree auto-detects the host (CUDA on Linux, Metal on macOS) and dispatches to make iree-cuda / make iree-metal, which run the validated setup scripts. make iree-env prints the resolved paths, pinned IREE version, and HAL drivers. All are idempotent: the scripts reuse an existing clone / build / venv, so re-running does not rebuild when the pinned artifact is present. A dedicated "IREE Backend Targets" group is added to make help.
• Reproducibility fix (both scripts). A runtime-only IREE configure at the pinned revision (e4a3b04) needs the benchmark and printf submodules: the top-level CMake pulls third_party/benchmark in under IREE_ENABLE_THREADING (on for the local-task driver) and third_party/printf under the runtime gates, neither gated by IREE_BUILD_*. The previous flatcc-only init therefore failed a fresh build at the configure/generate step. Both scripts now initialize every submodule except the heavy compiler-only ones (llvm-project, stablehlo, torch-mlir), which a runtime-only build (IREE_BUILD_COMPILER=OFF) never configures. The benchmark sentinel also self-heals an older flatcc-only clone.
• --info mode. Each script gains a read-only --info that prints the resolved paths, pinned version, and HAL drivers (surfaced by make iree-env); the eval-safe --env output is unchanged.
• README. The mlxcel-xla CUDA and macOS build sections now reference make iree instead of the ad-hoc clone / cmake / script invocation.

Validation (GB10, CUDA)

• make iree-env prints backend, home, iree-compile, CUDA root, pinned version, and HAL drivers.
• Reproducible from a clean source tree: removing the cloned src/ + build/ and running make iree-cuda re-clones, initializes flatcc + benchmark + printf, configures, and rebuilds libiree_runtime_unified.a (exit 0).
• cargo build --release --features xla-iree links the freshly built runtime.
• Token-exact on the GPU: greedy generation on Llama-3.2-1B (4-bit) is byte-identical between cuda f16 (default) and cuda f32 (48/48 tokens); f16 6.8 tok/s vs f32 4.85 tok/s.

Notes

• The Metal targets (make iree-metal) follow the same CMake gates and pass bash -n, but are not validatable on this Linux host; Metal validation belongs to a macOS session (tracked under feat: extend the fused low-precision path to the Metal target #575).
• Orthogonal observation (not addressed here): on the source-built CUDA runtime, MLXCEL_XLA_DEVICE=local-task reports "no driver 'local-task' registered" even though the driver is compiled in (libiree_hal_drivers_local_task_task_driver.a). This is a shim/runtime-registration behavior independent of this change and may warrant a follow-up.

Closes #571
Part of #570