Support ONNX opset 20 and faithful tanh GELU lowering

[matthargett]

Support standard ONNX opsets 19/20 for implemented operators, including both GELU formulas. This replaces the initial fail-closed patch with tested tanh support.

• Keep default/none GELU mapped to exact WebNN gelu; lower approximate="tanh" to the specified mul/add/tanh composition. Promote FP16 intermediates to FP32, then cast the result back. Reject malformed attributes and unsupported types.
• Preserve scalar inputs, avoid generated-name collisions, and emit WebNN's positional cast signature. Validate unsupported types before file/CLI constant folding.
• Audit opset changes: preserve AveragePool dilation and constant reduction axes, keepdims=0 and empty-axis semantics. Unsupported operators and dynamic axes remain explicit errors; accepting opset 20 does not imply complete ONNX coverage.

Validation:

• 216 tests, formatting and strict all-target Clippy pass on Rust 1.97 and 1.98. Tests exercise actual file import with optimization on/off and generated JavaScript calls through a numerical builder double.
• M4 CoreML: 72/72 tanh predictions pass against ONNX Runtime references through file import, graph conversion and typed dispatch. FP16/FP32, scalars, seeded tensors, 10,001-point sweeps, large tails and special values; CPU/GPU/NPU hints tested separately, without claiming hardware placement. Maximum finite absolute errors: FP32 9.54e-7, FP16 6.11e-5.

ONNX CTS already defines both formulas; WebNN WPT does not import ONNX. These regressions cover the missing importer boundary. Separately, native exact FP16 GELU under the NPU hint fails upstream ULP limits on this M4/macOS beta; that existing CoreML/harness issue is not a failure of the tanh decomposition and remains tracked in #257.