Add dataset quality checks

[zhiltsov-max]

When there is a dataset, it is often important to be able to check its quality. There are few way to check it:

• compare annotations against a specification. Find too small boxes, too large boxes, too close points etc.
• compare annotations from several sources, match by distance
• annotate a dataset with a model, compare annotations against the inference results

It would be nice to support these and other approaches to quality checking.

Currently, in 0.1.6, the approaches 2 and 3 are implemented as datum merge.

[zhiltsov-max]

Added dataset validation for classification, detection and segmentation tasks in #160, #227, #233