tSNE Updates

[alexcarssuncao]

sits_tsne: expose perplexity/rounds, add LSTM-FCN & LightTAE support, fix plot palette

Summary

This PR improves the t-SNE workflow and model coverage in sits:

• Parameters surfaced: perplexity and rounds are now explicit in sits_tsne() and stored in the result for reproducibility and plotting.
• New model support: sits_tsne() now works with LSTM-FCN and the current sits_lighttae implementation.
• Palette fix: Corrects the plot palette used by plot.sits_tsne().

Changes

1. Add perplexity and rounds parameters to sits_tsne() and propagate them to the returned object (used by plotting/metadata).
2. Add support for LSTM-FCN embeddings.
3. Update compatibility with the current sits_lighttae version.
4. Introduce model-identifying attributes/classes (e.g., "torch_model_mlp", "torch_model_tae", …) so sits_tsne() can route to the correct method.
5. Fix color palette selection in plot.sits_tsne() so class colors match SITS conventions.

Implementation notes

• sits_tsne() dispatches based on the model type; for now, I added attributes/classes to identify each torch model family.

Open question for maintainers

If there’s a preferred way to discover the concrete model without adding attributes, I’m happy to refactor. Possible alternatives:

• Use an S3 class vector on the model object (e.g., class(model) <- c("torch_model_tae", "torch_model")) and keep sits_tsne.torch_model_tae() etc.
• Add a small internal generic (e.g., .get_embeddings(model, x)) implemented per model type; sits_tsne() would call that.

Testing

• Added/updated tests to verify:
  • Structure of the returned sits_tsne object across models.
  • Perplexity clamping (too large / negative) still returns a valid object.

EXPERIMENTAL

• Attempted to fix the cpp issue in smooth.cpp:
  • Changed arma::is_finite() by std::infinite()

I’d appreciate any comments or suggestions—especially on the preferred mechanism for model-type dispatch (attribute vs. S3 class vector vs. helper generic).