Incremental versioning as complement to semver

When we have changing data (e.g. scheduled training jobs on application data) and want to snapshot it with TFDS into a nice, workable and highly performant format for model training, we currently need to bump the semver tag manually.

Could we introduce an alternative versioning scheme with auto-incrementing per execution of tfds build?

E.g. let dataset builders configure something like tfds.core.Version(tfds.core.VersioningScheme.AUTOINCREMENT) such that tfds build my_dataset writes artifacts to data_dir/my_dataset/config/1 and a subsequent tfds build my_dataset writes artifacts to data_dir/my_dataset/config/2

I could try a proposal PR if this sounds acceptable. I also understand if you'd rather stay committed to semver, of course, but having auto-incrementing versioning could make TFDS easier to integrate with TFX so it might be worth pursuing (semi-related https://github.com/tensorflow/datasets/issues/2082).

Start by reviewing the existing tfds.core.Version entry point and the behavior of tfds build my_dataset, then read the related issue 2082. Done would require an agreed versioning design and a proposal PR covering automatic per-build artifact versions and compatibility with existing semver behavior.