Fix deeplearning autoencoder implementation issues

[sergioald]

This PR fixes the implementation issues documented by the autoencoder smoke tests.

Changes include:

• Fix StandardAutoencoder and OrthogonalAutoencoder multidimensional flatten/reconstruction handling
• Add a ViT decoder projection from d_model tokens back to patch-dimension tokens before Unpatchify
• Make ConvLSTMAutoencoder reconstruct the last frame by default when y is not provided
• Make HybridConvLSTMTransformerAutoencoder reconstruct the last frame by default when y is not provided
• Store efficient_attention in the hybrid inner model
• Pass ConvLSTM kernel_size as a tuple
• Use deepcopy for best model state restoration

Local test result on Windows / Anaconda:
213 passed.

[sergioald]

Hi @tausiaj,

The CI checks have passed. This PR is ready for review.

It includes the autoencoder smoke tests and the implementation fixes discussed in the PR description. I also verified the full test suite locally on Windows/Anaconda with:

213 passed.

Happy to make any changes if you prefer a different design for the ConvLSTM/Hybrid reconstruction target or the StandardAutoencoder flattening behaviour.

After this PR, I would be happy to contribute follow-up improvements in separate PRs, for example reconstruction-error utilities, PCA/autoencoder comparison examples, or anomaly-detection examples for climate/ocean data.