Add initial ONNX importer crate with working op subset

[MagellaX]

Description

This introduces an initial ONNX import path encapsulated in a dedicated crate. The importer reads an ONNX model and produces a Luminal primitive graph plus named input/output GraphTensors for data ingress/egress. It is self-contained and does not affect existing runtime behavior.

Key points

• Provides a single entry point to load models and construct a runnable primitive graph, returning input/output tensors keyed by original ONNX names.
• Uses protobufs generated via prost with a vendored protoc, avoiding external build dependencies and reducing CI friction.
• Implements a pragmatic operator subset sufficient for basic models and initial end-to-end validation:
  • Constants and initializers
  • Elementwise Add/Sub/Mul/Div/Max/Min with simple broadcasting
  • Relu, Sigmoid, Tanh, Sqrt
  • MatMul and Gemm (alpha/beta/trans flags)
  • Softmax with axis (including negative indices)
  • Reshape (supports -1 and 0 when shape is constant), Transpose
  • Unsqueeze/Squeeze (via attribute or secondary input), Concat
• Converts ONNX shapes (including dim_param) into Luminal Expressions, mapping symbolic dims consistently.
• Marks graph outputs for retrieval to ensure values are retained after execution.
• Includes a smoke test that programmatically builds a small ONNX model (MatMul + Add + Softmax), imports it, runs the graph, and compares numerically to an equivalent Luminal graph.

Design notes

• The returned Graph is boxed to keep GraphTensor internal pointers stable while building and executing the graph.
• A trimmed ONNX proto is included to cover the importer’s current needs; it can be swapped or expanded as coverage grows.

Limitations and scope

• Not yet covering convolution, pooling, batch norm, gather/scatter, pad/clip, shape ops, or control flow.
• Focused on floating point tensors and common inference paths; quantized/complex dtypes are not handled.
• Broadcasting is intentionally simple and may need to be generalized to full ONNX semantics.
• Only import is implemented; export is out of scope here.

Near-term follow-ups

• Expand op coverage: Conv/ConvTranspose, pooling ops, Reduce ops (Mean/Sum), ArgMax, Pad/Clip, Flatten/Shape/Gather.
• Improve broadcasting to faithfully match ONNX rules across ranks and singleton dims.
• Extend dtype support (int/bool/bfloat16/float16) and test mixed-type behavior.
• Add an example CLI to load an ONNX file, feed inputs, and dump outputs for quick local validation.
• Add model-based tests (small public ONNX models) to exercise importer paths beyond the smoke test.
• Consider switching to the full upstream ONNX proto and documenting version compatibility.

Tracking

• Addresses the core of ONNX import requested in ONNX Integration #32 and complements the StableHLO path discussed in StableHLO support #85.