onnxruntime

48102b53 - [WebGPU] Keep InstanceNormalization scale and shift in f32

[WebGPU] Keep InstanceNormalization scale and shift in f32 WebNN conformance instance-normalization-zero-epsilon (float16 subnormal inputs, epsilon 0) returned Inf: inv_std is 2^24, which overflows when the per-channel scale/shift tensor is created with the f16 input type. Create that tensor as f32 and apply it in f32 on both the NCHW and NHWC paths, converting to the output type once. Tests use the WebNN vectors in f32 and f16, with and without affine, for both EP layouts. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>