Add iTransformer and pre-trained encoder-decoder transformer model providers

[nazim117]

Adds two new forecasting model backends and the factory wiring to select between them at runtime.

• iTransformer: inverted-attention transformer trained from scratch per asset; test coverage included.
• Encoder-decoder transformer: loads pre-trained weights from our HuggingFace repo; vendored network definition kept local to avoid runtime dependency on HF src layout.
• model_provider_factory.py: routes model_type config value to the correct provider.
• model_type.py / model_config.py: extended with new enum values and config fields for both models.
• Frontend: config editor updated to expose new model options.
• Docker: image updated to include new dependencies.