Improve embedding nets

[manuelgloeckler]

🚀 Feature Request: Pre-configured Embedding Networks for Common Data Types + General data shape support

It would be beneficial to have a suite of pre-configured embedding networks tailored for different data types readily available.

Currently, we provide a set of general methods, which require configuration, or leave it up to the user to choose or develop an appropriate embedding network. For common data types in SBI, it would be great to have several pre-configured networks that require minimal fine-tuning (e.g., users may only need to provide input shapes).

Potential architectures to include:

• ResNets - For fixed-size inputs
• 1D CNNs - Suitable for time-series data (optionally with temporal convolution)
• RNNs - Designed for time-series data
• Transformers - For permutation-invariant/general tasks i.e. see Add Transformer as embedding net #1413
• ...

We would like to have a general and "dynamic" data shape support (if combined with a suitable embedding network). There are currently a few redundant shape checks in our pipelines, which should be removed (#1324).

📌 Additional Context

If you identify a reasonable and general use case for a specific architecture and have the expertise to provide default configurations, feel free to create a sub-issue for the network you'd like to implement and begin working on it.

Note: The goal is not to include highly domain-specific architectures that are difficult to maintain.