Add @stdlib/onnx module for local ONNX Runtime inference

[nbeerbower]

Adds a new standard library module that enables running ONNX models
locally via libonnxruntime FFI, following the same pattern as the
sqlite and vector modules.

Architecture:

• stdlib/c/onnx_wrapper.c: Thin C wrapper that flattens OrtApi's
  struct-of-function-pointers into flat exported C functions callable
  through Hemlock's FFI system
• stdlib/onnx.hml: Pure Hemlock module providing high-level API for
  session management, tensor creation/inspection, and inference
• stdlib/docs/onnx.md: Full documentation with examples

API surface:

• Session lifecycle: create_session(), free_session()
• Model inspection: input_names(), output_names(), input_shape(),
  output_shape(), input_type(), output_type(), describe()
• Tensor creation: create_tensor_f32/f64/i32/i64(), free_tensor()
• Tensor inspection: tensor_shape(), tensor_data_f32/f64/i32/i64(),
  tensor_element_count(), tensor_type()
• Inference: run() with named (object) or positional (array) inputs
• Utilities: version(), api_version(), dtype_name()

Also includes:

• Makefile updates for conditional ONNX Runtime detection and wrapper
  compilation on Linux and macOS
• Test runner skip logic for when libonnxruntime is not installed
• Tests for constants, tensor operations, error handling, and inference

https://claude.ai/code/session_01PAARTAYAKLvdGGuiNaW9xF