Add mixed datasets, model truncation, new activation buffers

[adamkarvonen]

I made a few backward compatible changes here:

• First, I was training some Qwen 32B SAEs. I added a truncate_model() function, which deletes unneeded layers, which provides very significant memory savings on large models.
• I added an optional Pytorch activation buffer, which was easier to use with truncate_model(). I also added an Activault streaming activation buffer, which streams activations from S3.
• I added a data mixture dataset generator, which enables training on a mixture of 2 datasets. This is very useful for e.g. training on a mixture of pretrain and chat data.
• I also added an optional backup_steps, which saves the SAEs (state dict and optimizer) every x steps. This is useful when working with larger models, where training runs can be >24 hours.

I checked that the end to end test passed before and after.