[hiyufan]

The Relax ONNX frontend rejects HannWindow, HammingWindow and BlackmanWindow models. Add shared cosine-sum lowering using existing Relax operators and the existing tensor-to-shape helper for runtime lengths.

Support int32/int64 size inputs, initializer/runtime lengths, periodic/symmetric modes, the default float32 output and numeric output casts. Compute in float32, use ONNX's Hamming coefficients, and add the constant offset after the cosine terms to match native window-kernel rounding before integer casts. Validate the periodic attribute and positive initializer lengths.

Related to #20124 (these three window operators only).

Validation on main e345029c, Python 3.11, LLVM 21.1.8, ONNX 1.23.2 and ONNX Runtime 1.31.0:

• Before the converters: all 30 selected checker-valid cases fail to import, covering all three operators.
• All 120 added window tests passed across the matrix and dynamic-VM reuse runs. Coverage includes default/explicit periodic mode, initializer/runtime sizes, int32/int64 size dtypes, float16/float32/float64/int32 outputs, bfloat16 casts, lengths 1/2/7, invalid initializer lengths and invalid periodic values. Six reuse tests run the same compiled VM at lengths 1, 2, 7 and 10.
• Values, shapes and dtypes are checked against ONNX Runtime. Bfloat16 outputs use its float32 result followed by an independent NumPy bfloat16 cast.
• Adjacent Range tests: 11 passed; two float32 count tests return 32 elements versus ONNX Runtime's 33. Both failures reproduce on an untouched main worktree with the same native build and dependencies: test_range_constant_float and test_range_dynamic_scalar_inputs[-7.8--1.4-0.2-1-float32]. The Range converter is unchanged.
• File-scoped Ruff, ASF header/file-type and whitespace checks passed.

Prepared with OpenAI Codex assistance. The LLVM build and validation above were executed locally using the pinned TIRx/FFI sources.