Fix protein Conv1d dimensions and align CNN architecture with paper

[matija-marijan]

This pull request addresses the issues explained in #20, #12, and #13.

Changes:

• Apply Conv1d along protein sequence dimension instead of the embedding dimension (the protein embedding matrix is permuted, and the number of Conv1d input/output channels is updated)
• Align the protein convolution network architecture with the paper (three consecutive Conv1d layers).

Notes:

• The Conv1d dimension modification preserves sequential information in the protein embeddings.
• The three-layer protein CNN uses ascending channel sizes, with ReLU activations after each layer, following the DeepDTA implementation.