Fix intermediate frames trimming to make sure final notes output is not truncated

[hyperc54]

This PR fixes the issue highlighted in a different PR by @avan06 (thanks!).

The original lines

n_output_frames_original = int(np.floor(audio_original_length * (ANNOTATIONS_FPS / AUDIO_SAMPLE_RATE)))
    return unwrapped_output[:n_output_frames_original, :]  # trim to original audio length

trim the output frames a bit more than appropriate.

The root cause is that ANNOTATIONS_FPS, which represents the number of expected output frames per second, is rounded down to the next integer, dropping the few samples of the last incomplete frame. This makes audio_original_length * (ANNOTATIONS_FPS / AUDIO_SAMPLE_RATE)) underestimates the number of frames needed in the output (not much per frame, but it compounds to several seconds of audio when analysing audios that are >10mn long)

The PR adds a test to make sure that the last frame second (as computed downstream in the pipeline) is close to the actual audio length. (additionally, it would be nice to add a test for the model_frames_to_time method)

[avan06]

Thank you @hyperc54 and the team for identifying the root cause and fixing it!