Export to onnx fails for some models with Pytorch 1.12

[ORippler]

Describe the bug

in Pytorch 1.12, export to onnx fails for all models that rely on kornia.filters.gaussianblur2d.

To Reproduce

Steps to reproduce the behavior:

1. install PyTorch 1.12
2. run a model (e.g. padim) with export to onnx enabled

Expected behavior

Export to onnx works also in Pytorch 1.12

Additional context

I believe the error is due to missing shape annotations of torch.nn.functional.interpolate, at least that's where shapes are lost when inspecting via netron.

The problem can be fixed by reimplementing GaussianBlur to not rely on the input shape for computing the kernel (as a drawback num_channels have to be provided explicitly).