onnx-runtime-eager

Eager (single-op) execution for the ORT 2.0 runtime (docs/EAGER.md). Dispatch individual ONNX ops to execution-provider kernels without building a graph — a PyTorch-style experience with ONNX op semantics.

This crate is the pure-Rust core of the design. It reuses the existing runtime abstractions rather than inventing parallel ones:

• ExecutionProvider / Kernel / OpRegistry from onnx-runtime-ep-api,
• the populated CPU registry from onnx-runtime-ep-cpu,
• per-op shape/dtype inference from onnx-runtime-shape-inference,
• the IR vocabulary (Node, DataType, Attribute, …) from onnx-runtime-ir.

Phase-1 scope (docs/EAGER.md)

Implemented: [EagerContext], [EagerContext::dispatch] (single-op CPU dispatch), the DomainRegistry, opset resolution ([opset]), and the compiled-KernelCache.

DEFERRED (each is flagged at its hook point):

• PyO3 / Python bindings (§11), including the Tensor single→tensor / multi→tuple sugar and ops.* typed wrappers (§4.1).
• Subgraph ops If/Loop/Scan (§7).
• The opset context-manager / nxrt.device() context (§5.3, §2.2).
• GPU/CUDA EP dispatch and implicit cross-device transfer.
• Kernel-provided shape-inference fallback (§9.2).
• DLPack / numpy interop (§3).