pad cuts by num_samples as default

fix duration precision issue:

https://github.com/lhotse-speech/lhotse/blob/509f4a97ab79a9eaeb8ba377d2a8e4acc794e45b/lhotse/cut/set.py#L2967

0.15052083333333321 * 24000 = 3,612.5 = 3,613
0.15052083 * 24000 = 3,612.49992 = 3,612

lhotse/dataset/collation.py

@@ -189,7 +189,7 @@ def collate_audio(
 
     cuts = maybe_pad(
         cuts,
-        duration=max(cut.duration for cut in cuts),
+        num_samples=max(cut_id2num_samples.values()),
         direction=pad_direction,
         preserve_id=True,
     )