Here's a detailed description of all the changes in this PR:

Update ESC50 recipe: Add LMAC-TD support

This PR enhances the ESC50 recipe with LMAC-TD (Time Domain) interpretation capabilities and modernizes the data preparation script.

Changes to esc50_prepare.py:

• Updated repository URL to point to the current ESC-50 repository (karoldvl/ESC-50)
• Replaced custom SpeechBrain logger with standard Python logging
• Simplified data fetching by removing unnecessary parameters
• Fixed file path handling in archive extraction
• Cleaned up JSON file writing

Changes to interpret/README.md:

• Added comprehensive LMAC-TD documentation section
• Included performance metrics table with LMAC-TD results across different alpha values
• Added training instructions for LMAC-TD with ESC50 dataset
• Documented WHAM! noise augmentation option for improved interpretations
• Added citation and references to LMAC-TD ICASSP paper and companion website

Changes to interpret/eval.py:

• Added support for LMAC-TD evaluation alongside existing LMAC and L2I methods
• Implemented conditional model loading based on interpretation method
• Added LMAC-TD-specific single sample processing with time-domain output
• Updated import statements and class references for LMAC-TD
• Enhanced evaluation workflow to handle different interpreter architectures

Changes to interpret/hparams/lmactd_cnn14.yaml:

• Created LMAC-TD specific hyperparameters configuration with settings presented in the paper
• Updated experiment name and method settings for LMAC-TD
• Optimized model architecture parameters (reduced layers, heads, FFN dimensions)
• Updated STFT parameters (n_fft increased to 2048)
• Note for users with limited VRAM: You can reduce memory usage by decreasing N_encoder_out, out_channels, and d_ffn parameters of the SBTransformerBlocks

Changes to interpret/train_lmactd.py:

• Added LMACTD class
• Updated training recipe comments and documentation for LMAC-TD
• Simplified mask processing
• Enhanced computation steps for time-domain interpretation
• Updated author information and maintained compatibility with existing workflow

New additions to speechbrain/lobes/models/Cnn14.py:

• Added CNN14PSI_encoderdecoder class
• Implemented progressive channel reduction architecture
• Provided comprehensive documentation and usage examples

Updates to speechbrain/lobes/models/dual_path.py:

• Added configurable activation function support to Dual_Path_Model
• Implemented sigmoid activation option alongside existing ReLU
• Enhanced model flexibility for different interpretation methods

File cleanup:

• Removed obsolete lmac_sepformerstuff.py file (1,700 lines deleted)

These changes add new time-domain interpretation capabilities while ensuring compatibility with the current ESC-50 dataset repository and improving code maintainability. The LMAC-TD method provides enhanced interpretability for audio classifiers with direct time-domain explanations.

@fpaissan could you update the PR description with this and mark as ready for review when the checks pass?