Align audio array shape

[emanuele-moscato]

Various fixes:

• Align the audio array reshaping with what librosa needs for resampling (when working with mono audio, librosa requires the audio array to have shape (n_samples,), not (n_samples, 1) as we have).
• Fix bug related to input features to speech models (sometimes the feature extractor returns input tensors under the input_values key, sometimes under the input_features one). Note: should this be done for all the model helpers?
• In utils_removal.py, fix the case in which the offset a subtracted from the word start timestamp would have caused the audio to be read from the end.
• In utils_removal.py, fix the paths for the (local) mp3 files with white and pink noise.
• Added tests for text and speech
• Readme updated and Testing.md created

To do after merging the testing_environment branch:

• Make sure that tests still work.
• Add pytest among the requirements (pyproject.toml).
• Try creating a new environment from scratch, install all the requirements (watch out for the manual ones!) and make sure tests still run.