Implement missing preprocessing and postprocessing (per sample)

[m-novikov]

Main issue about implementing pre/post-processing in tiktorch.
For now, we decided to go only with per_sample implementation per_dataset would be added after
Descriptions in configuration spec

• binarize
• clip
• scale_linear
• sigmoid
• zero_mean_unit_variance partially completed

Finished implementation should contain:

1. Snapshot tests
2. Should correctly handle axes keyword argument if present in the spec
3. Should have a docstring
4. Should use xarray/numpy for implementation and not framework-specific tools
5. Should correctly propagate axis tags

[k-dominik]

Should scale_range be added here, too? or can it already be found some place else?