onnxruntime

9c21c823 - [WebGPU] Fix Pow scalar operands in vectorized broadcasting (#33348)

[WebGPU] Fix Pow scalar operands in vectorized broadcasting (#33348) ### Description Pow can load a scalar operand in its vectorized broadcast path, including base [3,2,64,64] with exponent [1,1,1,1], then pass it to the vec4-only pow_v helper. Convert each argument to vec4 of its own element type so scalar loads broadcast correctly and mixed base/exponent types are preserved. Add six WebGPU regressions covering the reported shape, broadcast in both directions, row-wise mixed float/int32, and FP16, each with dynamic and initializer exponents and CPU fallback disabled. Validation on Windows x64 / native WebGPU / D3D12: - All six new tests fail before the fix and pass afterward: 12 confirmed GPU executions, zero skips. - Existing Pow tests: 18 reported passes, zero GTest skips; 11 GPU executions. Six double/int64 cases are CPU-only, and one generic FP16 helper is inactive without CUDA/CoreML. - Fresh and incremental builds, changed-line clang-format, and git diff --check passed. Metal and browser/Wasm were not tested. ### Motivation and Context Fixes #32842.