Optimization: Cash kernels and move creation to correct device

[zakajd]

Is your feature request related to a problem? Please describe.
Most metrics use some kind of kernels for extraction of image features.
Now those kernels are created each time when metric is called. That's slow and redundant when the function is called hundreds of times.

Another problem related to performance is creation of temporal tensors first on CPU and later moving them to GPU if needed (.to(x.device)).

Describe the solution you'd like

1. Add support for cashing kernels between runs. Add kernel param to functional API. Create and store kernel in class API
2. Replace .to(x.device) and simmular calls to explicit creation of tensor on target device. Like torch.ones(N, N, device=x.device).

Additional context
For metrics that use more than one kernel implementation details can be discussed.

[denproc]

Hi @zakajd,

The second part of the proposed solution is partially implemented in #334.

While the proposed changes cover the tensor creation related to filters and colour conversions, some parts of the library could still use .to (piq/base.py, fid.py, perceptual.py, pieapp.py and piq/functional/resize.py).