Fix FP16 conversion for initializer-backed graph outputs

[Qiong Wu (qiowu) (DingmaomaoBJTU)]

Summary

• handle safe top-level FP32 graph outputs supplied directly by dense initializers during FP16 conversion
• preserve exact FP32 initializer/output values when keep_io_types=True
• convert the backing initializer data when producing pure FP16 I/O
• reject shared, nested, overridable, unloaded-external-data, or generated-name-collision cases before ORT mutates the model

Root cause

ORT's FP16 converter assumes every graph output has a node producer. For an output supplied directly by an initializer:

• keep_io_types=True creates an orphan graph_output_castN input, and WinML topological sorting fails with Graph is not a DAG
• keep_io_types=False changes the graph-output declaration to FLOAT16 but leaves the backing initializer FLOAT, causing strict type inference / ORT session creation to fail

The fix captures only provably safe top-level initializer outputs. It removes the exact orphan Cast for preserved FP32 I/O, or converts the resident initializer data for pure FP16. Ambiguous mixed-precision semantics fail closed.

Validation

• pytest tests/unit/optim/test_fp16.py -q — 27 passed
• related optimization/quantization suite — 672 passed, 16 skipped, 1 xfailed
• ruff check src/winml/modelkit/quant/fp16.py tests/unit/optim/test_fp16.py
• mypy -p winml.modelkit — 429 source files clean
• public winml quantize --precision fp16 reproduction on keen_hominy now succeeds; the generated model creates an ORT session with all 7 outputs, while the initializer-backed landmarks_2d output remains FP32 under the default keep-I/O policy

Scope

This intentionally does not guess semantics for initializer outputs that are shared with graph consumers, captured by nested graphs, exposed as overridable graph inputs, or backed by unloaded external data. Those cases return an explicit error before conversion rather than publishing an invalid model.

[xieofxie]

One correctness issue remains in duplicate-initializer validation for subgraphs beneath blocked operators.

[xieofxie]

Two correctness gaps remain around nested initializer outputs and lexical consumer scope.

[xieofxie]

Two conversion regressions remain for overridable initializer outputs and sparse FLOAT initializers.

[xieofxie]

One correctness issue remains in keep-I/O handling for nested graph input shadowing.

[xieofxie]

One correctness issue remains in nested keep-I/O collision detection.

[xieofxie]

All previously identified correctness issues have been addressed. The latest changes look good to merge.