[Feature Request] [web/webgpu] Support non-symmetrical padding in Conv

[poke1024]

Describe the feature request

ONNX runtime web with ep:webgpu currently expects padding in Conv to be totally symmetrical:

onnxruntime/js/web/lib/wasm/jsep/webgpu/ops/3rd-party/conv3d_naive_webgpu.ts

Lines 125 to 127 in 9681fe6

	if (!pad.every((val, _, arr) => val === arr[0])) {
	throw Error(`Unsupported padding parameter: ${pad}`);
	}

Models that need Conv with asymmetrical padding fail to run with the error shown above.

A possible workaround for users might involve factoring out the padding into a separate Pad op. This is tricky to do for models with pretrained weight however.

Describe scenario use case

Necessary for executing any model that needs convolutions with Conv with non-symmetric padding under ONNX runtime web with ep:webgpu. One example of such a model is TransnetV2 (https://github.com/soCzech/TransNetV2).

[fs-eire]

I think WebGPU EP should work for this scenario and we are working on a build to use webgpu ep with emscripten