More robust evaluation of assignments

[amyreese]

Improves assignment evaluation in a number of ways:

• Evaluates assignments in reverse line-number order, from bottom to top
• Tracks target name and line number when evaluating recursively
• Only considers recursive assignments that occur before the current target line
• Attempts evaluating earlier assignments if later assignments are inconclusive
• Prevent evaluating binary operands if either side evaluates to "??"
• Add support for "augmented assignments" like +=

This enables support for evaluating self-referential assignments like:

foo = foo + bar

or

foo += bar

Adds a basic test covering both self-referential assignment types.

Fixes #56

[amyreese]

Final metadata evaluated against onnx repo:

(.venv) jreese@mordin ~/workspace/dowsing onnx  » python -m dowsing.pep517 ~/scratch/onnxruntime | jq .get_metadata
{
  "metadata_version": "2.1",
  "name": "onnxruntime-openvino",
  "version": "??",
  "summary": "ONNX Runtime is a runtime accelerator for Machine Learning models",
  "description": "??",
  "keywords": "onnx machine learning",
  "home_page": "https://onnxruntime.ai",
  "author": "Microsoft Corporation",
  "author_email": "onnxruntime@microsoft.com",
  "license": "MIT License",
  "classifiers": [
    "Development Status :: 5 - Production/Stable",
    "Intended Audience :: Developers",
    "License :: OSI Approved :: MIT License",
    "Operating System :: POSIX :: Linux",
    "Topic :: Scientific/Engineering",
    "Topic :: Scientific/Engineering :: Mathematics",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
    "Topic :: Software Development",
    "Topic :: Software Development :: Libraries",
    "Topic :: Software Development :: Libraries :: Python Modules",
    "Programming Language :: Python",
    "Programming Language :: Python :: 3 :: Only",
    "Programming Language :: Python :: 3.6",
    "Programming Language :: Python :: 3.7",
    "Programming Language :: Python :: 3.8",
    "Programming Language :: Python :: 3.9"
  ],
  "requires_dist": "??",
  "package_data": "??",
  "packages": [
    "onnxruntime",
    "onnxruntime.backend",
    "onnxruntime.capi",
    "onnxruntime.capi.training",
    "onnxruntime.datasets",
    "onnxruntime.tools",
    "onnxruntime.tools.ort_format_model",
    "onnxruntime.tools.ort_format_model.ort_flatbuffers_py",
    "onnxruntime.tools.ort_format_model.ort_flatbuffers_py.fbs",
    "onnxruntime.quantization",
    "onnxruntime.quantization.operators",
    "onnxruntime.quantization.CalTableFlatBuffers",
    "onnxruntime.transformers",
    "onnxruntime.transformers.longformer"
  ],
  "packages_dict": {
    "onnxruntime": "onnxruntime",
    "onnxruntime.backend": "onnxruntime/backend",
    "onnxruntime.capi": "onnxruntime/capi",
    "onnxruntime.capi.training": "onnxruntime/capi/training",
    "onnxruntime.datasets": "onnxruntime/datasets",
    "onnxruntime.tools": "onnxruntime/tools",
    "onnxruntime.tools.ort_format_model": "onnxruntime/tools/ort_format_model",
    "onnxruntime.tools.ort_format_model.ort_flatbuffers_py": "onnxruntime/tools/ort_format_model/ort_flatbuffers_py",
    "onnxruntime.tools.ort_format_model.ort_flatbuffers_py.fbs": "onnxruntime/tools/ort_format_model/ort_flatbuffers_py/fbs",
    "onnxruntime.quantization": "onnxruntime/quantization",
    "onnxruntime.quantization.operators": "onnxruntime/quantization/operators",
    "onnxruntime.quantization.CalTableFlatBuffers": "onnxruntime/quantization/CalTableFlatBuffers",
    "onnxruntime.transformers": "onnxruntime/transformers",
    "onnxruntime.transformers.longformer": "onnxruntime/transformers/longformer"
  },
  "entry_points": {
    "console_scripts": [
      "onnxruntime_test = onnxruntime.tools.onnxruntime_test:main"
    ]
  }
}

[thatch]

A test would be nice, even if it's invalid python (unbound local x=x).

[amyreese]

Running this on onnxruntime now produces "version": "+cpu", which I don't like as much as "??" for this specific instance—but it's something I guess?—but would also correctly allow evaluating empty strings, where the original rev of this PR would have discarded x = '' and returned "??" instead.

[thatch]

Seems to produce issues when redefining builtins; see https://gist.github.com/thatch/ebc427075958a2338b7b7d712f77b01e for a simplified version of one that is copypasted over many popular projects' setup.py