pnnx: handle float16 scalar constant when importing onnx

[Lfan-ke]

The onnx importer's scalar-constant loader in pass_onnx.cpp handles INT32/INT64/FLOAT/BOOL but has no FLOAT16 branch, so a float16 scalar constant leaves params["value"] unset. Later pass_level3/eliminate_noop_math.cpp calls params.at("value") and throws std::out_of_range (map::at), aborting the conversion.

This adds a FLOAT16 branch that reuses the existing float16_to_float32 helper, reading either raw_data or the int32_data bit pattern, so float16 scalar constants are handled like the other scalar types.

It reproduces with a float16 GELU exported at opset 18 (dynamo), which lowers gelu into primitives whose 0.5 / sqrt(2) coefficients become float16 scalar constants. Added tests/onnx/test_F_gelu_fp16.py for it - the test aborts before this change and passes after.

Closes #6817

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 143e074fc9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Gate dynamo exporter test on torch versions

This test is now registered for every pnnx CTest run, including matrix jobs such as Python 3.9 with torch 1.11.0 in .github/workflows/pnnx.yml before the workflow runs ctest. In those older torch environments torch.onnx.export does not accept the dynamo exporter option, so the test fails with a TypeError before it ever reaches the importer; please skip/gate this case on a torch version that supports dynamo=True.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Avoid running fp16 GELU through ORT CPU

In the pnnx workflow the Python side installs the CPU onnxruntime package, so this session runs on CPU; the exported fp16 GELU decomposition uses fp16 primitive ops, but ONNX Runtime's CPU kernel table lists Erf for CPU as tensor(float) only (operator kernels). On torch versions new enough to export the model, this makes the reference inference fail in ORT instead of testing pnnx; use the PyTorch output as the reference, cast the ORT graph to fp32, or skip unless a float16-capable EP is available.

Useful? React with 👍 / 👎.

[Lfan-ke]

Both fixed: the test now gates on torch >= 2.5 (older matrix jobs without dynamo=True skip cleanly), and the reference is the torch model output instead of ORT (drops the ORT fp16-Erf dependency). Local pnnx-vs-torch max diff 5e-4, test passes.