Add multimodal RNN support

[jhnwu3]

This pull request introduces a new multimodal RNN model to the pyhealth library, enabling the handling of both sequential and non-sequential input features for clinical prediction tasks. It also provides a comprehensive example script for using the new model on the MIMIC-IV dataset for in-hospital mortality prediction. Additionally, the documentation and model API are updated to reflect these changes.

Major changes include:

New Model: MultimodalRNN

• Added the MultimodalRNN class to pyhealth.models.rnn, which automatically distinguishes between sequential and non-sequential features and processes them appropriately (sequential features via RNN layers, non-sequential features via direct embedding and pooling). The model concatenates all feature representations for final prediction.
• Exported MultimodalRNN in the pyhealth.models package init file for public use.
• Updated the API documentation to include MultimodalRNN.

Example Usage

• Added a detailed example script mortality_mimic4_multimodal_rnn.py demonstrating how to use the new MultimodalRNN model for mortality prediction with mixed feature types on the MIMIC-IV dataset. The script covers data loading, task setup, model training, evaluation, and sample predictions.

Improvements and Bug Fixes

• Improved masking in RNN forward passes by using the absolute value before summing, ensuring correct mask computation even if embeddings sum to zero due to negative values.
• Refactored imports and type annotations in pyhealth.models.rnn to support the new model and feature classification logic.

[Logiquo]

I think [0.0, 0.0] could still be a valid embedding valid even for non-padding token, though very unlikely. But this can be considered as a temporary fix for now i think.

We may want to add a TODO here.

[jhnwu3]

Yeah, I think we may need to change our EmbeddingModel to be able to return a pad/mask tensor itself through the use of our processor vocabulary, and have it as a used return option here.

[jhnwu3]

Because I don't think the mask here really does very much here lol the more I think about it. It just assumes that somehow a 0 embedding from the EmbeddingModel is properly assigning 0 values to the sequence embeddings, etc.