HYD-315 Feature/auto embeddings

[BeardedWhale]

Changes

• Added logic for creation of autoembeddings:
  • Added function that infers type of transformation for each column of training/production data using profile type of model input
  • Added AutoEmbeddingsEncoder class that creates embeddings for tabular data automatically using transformation type
  • Added saving/loading autoembeddigns encoder to/from S3
• Added UMAP transformer for mixed data types
• Refactored /supported endpoint logic. Added use cases for autoembeddings:
  • Model must training data if there is no 'embeddings' field
  • Check that model has >2 inputs with shape==scalar and profile!=[None, Image, Text]
• Splited transform_high_dimensional() into transform_high_dimensional() and transform_high_dimensional_mixed()
• Removed support for vis_metrics and nearest neighbours for mixed data type embeddings.

Now to make data projection work model should have one of the following properties:

• 'embeddings' field
• training data and >2 inputs with shape==scalar and profile!=[None, Image, Text]