Crop synthesized audio in fast_fluidsynth

Sometimes (maybe .1% of the time), fluidsynth would pad a bunch of
silence to the end of a synthesis.  This resulted in a CQT with a big
silent block, which would result in false matches to the 2 (+?) tracks
in the MSD whose 7digital preview clips were silent (they have hidden
tracks).  Probably causes issues elsewhere, too.

feature_extraction.py

@@ -49,6 +49,9 @@ def fast_fluidsynth(m, fs):
              temp_mid.name], stderr=devnull)
     # Load from temp wav file
     audio, _ = librosa.load(temp_wav.name, sr=fs)
+    # Occasionally, fluidsynth pads a lot of silence on the end, so here we
+    # crop to the length of the midi object
+    audio = audio[:int(m.get_end_time() * fs)]
     # Close/delete temp files
     temp_mid.close()
     temp_wav.close()