Create a fused 2× bilinear upsampling + 3×3 convolution + SiLU operator.
Baseline: perform upsampling per-channel with loops and run conv per-output-channel sequentially.
Optimized: use torch.interpolate once, channels_last tensors, and a single conv2d with fused SiLU.
Provide Model/ModelNew plus get_inputs/get_init_inputs, and a run_code identical in structure to the root run_code.

