Fix activation hook batching, padding trim, and decoding dimensionality

[anandawolz]

This PR fixes critical issues in the activation-extraction pipeline and downstream decoding. Previously, activation hooks overwrote outputs per batch, causing loss of data and shape mismatch. Additionally, concatenation and padding removal were not handled for batches >1.

Changes:

• Hook Accumulation: Changed the activation hook to accumulate outputs per mini-batch in a list.
• Trimming & Concatenation: Trim padding uniformly across batches. Concatenate along the batch dimension (axis=0) to form (n_windows, channels, time) arrays and then:
• Reshape: remove the batch dimension to yield (channels, time) arrays to avoid error for kNN decoding, ensuring compatibility with scikit-learn

tests are not passing because _cut_array doesn't handle 2D activation arrays yet. This is handled in PR #59

[CeliaBenquet]

lgtm! As discussed with @anandawolz, some of the changes are overridden and tests pass in #59. @anandawolz can you update the main comment with what was not functional in this one.

[CeliaBenquet]

Minor and no need to reverse, but in general it is better software engineer practice to create a variable in the if statement and return only once, once you exit the if-else: https://jdriven.com/blog/2021/10/Single-return-vs-Multiple-returns.