Fix gated forward functions

[callummcdougall]

2 fixes:

1. In the forward method, when self.use_error_term is True, for non-standard architectures things like centering and input activation functions weren't being applied correctly. I think this was just a mistake, and is now fixed.
2. In the encode_gated method, the hook hook_sae_acts_post pointed to the activations post-ReLU (or whatever activation function is), but not to the actual output of this function, i.e. post-ReLU activations multiplied by masking values. I think a long-term solution has 2 separate hooks for each of these (e.g. hook_sae_acts_post and hook_sae_mag_post), but if we just have a single hook called hook_sae_acts_post then I think it makes a lot more sense for it to refer to the output of the encoder.

[jbloomAus]

Thanks!

[callummcdougall]

np! also let me know if there are things I can do to not require formatting PRs in the future - would this involve something like using the same workspace config files as this repo does?

[jbloomAus]

@callummcdougall run make format and make check-ci nothing more to it!