Unified autoencoder loader

[SrGonao]

Have a unified way to load autoencoders into the model for caching. Should allow for some flexibility for people to add new autoencoders. Completed when:

• Function returns a hooked model for some choice of autoencoder
• Support for SAE lens
• Tests that verify that an autoencoder is correctly "hooked"

[luciaquirke]

Closing because the LatentCache now takes a hookpoint_to_sparse_encode dictionary that can be built for various autoencoders and other models as follows:

hookpoint_to_sparse_encode = {
   hookpoint: autoencoders[hookpoint].forward
   for hookpoint in hookpoints
}