Add option to split AMI segments similar to Kaldi

[desh2608]

Kaldi AMI data preparation includes splitting long segments based on full-stops and commas. This is useful for training ASR models since otherwise we would throw away a lot of data. Here are some statistics before and after adding this option:

# segments	Train	Dev	Test
Kaldi	105149	13059	12612
Lhotse (w/o splitting)	65557	8665	7490
Lhotse (w/ splitting)	103377	12614	12121

The remaining difference may be because we don't "over-segment" the segments if they are already under the max words.

[desh2608]

(@HuangZiliAndy and I have both observed that decoding WERs can be significantly better with the segment splitting.)

[pzelasko]

Great!