Refactor activation normalization handling and remove ActivationNormalizer class, such that autoloading of activationnormalizer with load_pretrained works automatically.

This commit includes the following changes:

- Removed the `ActivationNormalizer` class from `utils.py`, centralizing activation normalization logic within the `NormalizableMixin`.
- Updated the `NormalizableMixin` to accept mean and standard deviation tensors directly, replacing the previous reliance on `ActivationNormalizer`.
- Modified the constructors of `BatchTopKSAE`, `CrossCoder`, and their trainers to accept `activation_mean` and `activation_std` parameters instead of an `activation_normalizer`.
- Adjusted normalization and denormalization methods to utilize the new mean and std tensors, ensuring that activations are processed correctly.
- Cleaned up related code in `cache.py`, `dictionary.py`, and various trainer files to reflect these changes.

These modifications enhance the clarity and maintainability of the code while ensuring proper handling of activation normalization across various components.

Author: jkminder

dictionary_learning/cache.py

@@ -15,7 +15,6 @@
     dtype_to_str,
     str_to_dtype,
     torch_to_numpy_dtype,
-    ActivationNormalizer,
 )
 
 if DEBUG:
@@ -303,9 +302,6 @@ def std(self):
                 )
         return self._std
 
-    @property
-    def normalizer(self):
-        return ActivationNormalizer(self.mean, self.std)
 
     @property
     def running_stats(self):
@@ -701,9 +697,6 @@ def std(self):
             (self.activation_cache_1.std, self.activation_cache_2.std), dim=0
         )
 
-    @property
-    def normalizer(self):
-        return ActivationNormalizer(self.mean, self.std)
 
 
 class ActivationCacheTuple:
@@ -732,7 +725,3 @@ def mean(self):
     @property
     def std(self):
         return th.stack([cache.std for cache in self.activation_caches], dim=0)
-
-    @property
-    def normalizer(self):
-        return ActivationNormalizer(self.mean, self.std)