Add self-supervised time-series representation learning

**Is your feature request related to a problem? Please describe.**
No

**Describe the solution you'd like**
Add native support for learning feature embeddings from time-series data using self-supervised learning.

The learned representation should be usable as input to existing ML.NET trainers and support standard ML.NET model persistence and pipelines.

**Describe alternatives you've considered**
None

**Additional context**
Relevant implementations:

https://github.com/yuqinie98/PatchTST

https://github.com/yuezhihan/ts2vec

Start by reviewing ML.NET's existing trainer, model-persistence, and pipeline extension points, then compare the PatchTST and ts2vec implementations linked in the issue. Done means defining and implementing native time-series representation learning that produces embeddings usable by existing ML.NET trainers and persists through standard ML.NET pipelines.

Written by the indexing model from the issue text.