conv2d benchmark

Benchmarking fp32 conv2d kernels on 1x64x256x256 input with 128 filters of size 3x3.

• "naive" is a simple nested-loop WebGPU kernel
• "onnx" runs a Conv operator in onnxruntime-web
• "tfjs" runs tf.conv2d() with NHWC format
• "jax-js" runs jax.lax.convGeneralDilated()